/usr/local/lib/python3.12/dist-packages/torch/cuda/__init__.py:61: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
  import pynvml  # type: ignore[import]

Loading checkpoint shards:   0%|          | 0/5 [00:00<?, ?it/s]
Loading checkpoint shards:  20%|██        | 1/5 [00:05<00:20,  5.11s/it]
Loading checkpoint shards:  40%|████      | 2/5 [00:10<00:16,  5.43s/it]
Loading checkpoint shards:  60%|██████    | 3/5 [00:16<00:11,  5.54s/it]
Loading checkpoint shards:  80%|████████  | 4/5 [00:22<00:05,  5.74s/it]
Loading checkpoint shards: 100%|██████████| 5/5 [00:25<00:00,  4.82s/it]
Loading checkpoint shards: 100%|██████████| 5/5 [00:25<00:00,  5.13s/it]
model sequence length: 4096
exp time: 2026-02-23 08:07:13
args: VPTQArguments(model_name='/workspace/models/bielik-mistral-11b-instruct', seq_len=4096, quant_step=1, percdamp=0.01, blocksize=128, output_dir='/workspace/variant-e/output/2026-02-23-08-06-48', seed=0, eval=False, new_eval=True, save_model=True, save_packed_model=True, disable_actorder=False, hessian_path='/workspace/hessians/quip-format/hessians', inv_hessian_path=None, num_gpus=1, eval_nsamples=128, save_qlinear=False, absorb_perm=False)
quant_args: QuantizationArguments(vector_lens=[-1, 8], num_centroids=[-1, 65536], num_res_centroids=[-1, 256], npercent=0.0, group_num=1, group_size=-1, kiter=100, ktol=1e-05, kseed=0, kmeans_mode='hessian', kmeans_alpha=0, enable_norm=True, norm_dim=0, enable_perm=True)
Starting VPTQ...
model dtype: torch.bfloat16
----quantization start ...---- 2026-02-23 08:07:15
gpu 0 tasks: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49]
/usr/local/lib/python3.12/dist-packages/torch/cuda/__init__.py:61: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
  import pynvml  # type: ignore[import]
INFO - ----Quantizing on cuda:0----
INFO - ----Quantizing layer 0 ...---- 2026-02-23 08:07:19 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/0_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:07:20 0.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:0.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 1509.1005859375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 0.self_attn.q_proj 1st kmeans time: 38.29729413986206
INFO - 0.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 0.self_attn.q_proj proxy error before VPTQ: 4.021879362881009e-07, 4.613467899616808e-05, 0.008717692457139492
INFO - 0.self_attn.q_proj 1st error time: 2.3960700035095215
INFO - 0.self_attn.q_proj proxy error after VPTQ: 3.4297522688575555e-07, 4.613467899616808e-05, 0.007434217259287834
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 90 iterations, error 747.7081298828125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 0.self_attn.q_proj residual time: 0.2952842712402344
INFO - 0.self_attn.q_proj 2ed gptq time: 2.6295106410980225
INFO - 0.self_attn.q_proj 2ed error time: 0.0009756088256835938
INFO - 0.self_attn.q_proj proxy error after residual VPTQ: 1.1542931588337524e-07, 4.613467899616808e-05, 0.002502007642760873
INFO - load Hessian from /workspace/hessians/quip-format/hessians/0_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:08:05 0.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:0.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 36 iterations, error 371.0067138671875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 0.self_attn.k_proj 1st kmeans time: 3.6074445247650146
INFO - 0.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 0.self_attn.k_proj proxy error before VPTQ: 5.77489487341154e-08, 2.476580993970856e-05, 0.002331801224499941
INFO - 0.self_attn.k_proj 1st error time: 1.501157522201538
INFO - 0.self_attn.k_proj proxy error after VPTQ: 4.7689283633189916e-08, 2.476580993970856e-05, 0.0019256096566095948
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 96 iterations, error 168.75799560546875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 0.self_attn.k_proj residual time: 0.08601808547973633
INFO - 0.self_attn.k_proj 2ed gptq time: 1.9622139930725098
INFO - 0.self_attn.k_proj 2ed error time: 0.0003750324249267578
INFO - 0.self_attn.k_proj proxy error after residual VPTQ: 1.512740865905471e-08, 2.476580993970856e-05, 0.0006108182715252042
INFO - load Hessian from /workspace/hessians/quip-format/hessians/0_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:08:14 0.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:0.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 21 iterations, error 1798.6630859375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 0.self_attn.v_proj 1st kmeans time: 2.183669090270996
INFO - 0.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 0.self_attn.v_proj proxy error before VPTQ: 2.118130026929066e-08, 7.023785201454302e-06, 0.0030156532302498817
INFO - 0.self_attn.v_proj 1st error time: 1.5029101371765137
INFO - 0.self_attn.v_proj proxy error after VPTQ: 2.2339094130074955e-08, 7.023785201454302e-06, 0.003180492203682661
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 74 iterations, error 396.33184814453125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 0.self_attn.v_proj residual time: 0.07417798042297363
INFO - 0.self_attn.v_proj 2ed gptq time: 1.9461991786956787
INFO - 0.self_attn.v_proj 2ed error time: 0.00034737586975097656
INFO - 0.self_attn.v_proj proxy error after residual VPTQ: 3.963925188088524e-09, 7.023785201454302e-06, 0.0005643573822453618
INFO - load Hessian from /workspace/hessians/quip-format/hessians/0_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:08:21 0.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:0.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 36 iterations, error 6393.97998046875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 0.self_attn.o_proj 1st kmeans time: 14.27430772781372
INFO - 0.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 0.self_attn.o_proj proxy error before VPTQ: 4.38875247255055e-07, 0.0011774187441915274, 0.000372743554180488
INFO - 0.self_attn.o_proj 1st error time: 2.2940142154693604
INFO - 0.self_attn.o_proj proxy error after VPTQ: 2.7321146944814245e-07, 0.0011774187441915274, 0.00023204274475574493
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 2860.195068359375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 0.self_attn.o_proj residual time: 0.3076467514038086
INFO - 0.self_attn.o_proj 2ed gptq time: 2.5574703216552734
INFO - 0.self_attn.o_proj 2ed error time: 0.0009515285491943359
INFO - 0.self_attn.o_proj proxy error after residual VPTQ: 8.912078186540384e-08, 0.0011774187441915274, 7.569166336907074e-05
INFO - load Hessian from /workspace/hessians/quip-format/hessians/0_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:08:42 0.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:0.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 61480.609375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 0.mlp.gate_proj 1st kmeans time: 132.60115122795105
INFO - 0.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 0.mlp.gate_proj proxy error before VPTQ: 1.1984515367657878e-05, 0.001366345677524805, 0.008771217428147793
INFO - 0.mlp.gate_proj 1st error time: 4.8356194496154785
INFO - 0.mlp.gate_proj proxy error after VPTQ: 5.531853730644798e-06, 0.001366345677524805, 0.004048648755997419
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 26975.1328125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 0.mlp.gate_proj residual time: 1.0569195747375488
INFO - 0.mlp.gate_proj 2ed gptq time: 5.083466053009033
INFO - 0.mlp.gate_proj 2ed error time: 0.001214742660522461
INFO - 0.mlp.gate_proj proxy error after residual VPTQ: 1.6522285477549303e-06, 0.001366345677524805, 0.0012092316756024957
INFO - load Hessian from /workspace/hessians/quip-format/hessians/0_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:11:08 0.mlp.up_proj
INFO - enabling norm dim 0, layer_name:0.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 67553.890625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 0.mlp.up_proj 1st kmeans time: 132.41159057617188
INFO - 0.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 0.mlp.up_proj proxy error before VPTQ: 8.159808203345165e-06, 0.0002270069089718163, 0.03594519942998886
INFO - 0.mlp.up_proj 1st error time: 4.788538932800293
INFO - 0.mlp.up_proj proxy error after VPTQ: 2.972488573504961e-06, 0.0002270069089718163, 0.013094265013933182
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 28407.037109375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 0.mlp.up_proj residual time: 1.0042574405670166
INFO - 0.mlp.up_proj 2ed gptq time: 5.025835990905762
INFO - 0.mlp.up_proj 2ed error time: 0.0014598369598388672
INFO - 0.mlp.up_proj proxy error after residual VPTQ: 8.649269602756249e-07, 0.0002270069089718163, 0.0038101349491626024
INFO - load Hessian from /workspace/hessians/quip-format/hessians/0_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:13:36 0.mlp.down_proj
INFO - enabling norm dim 0, layer_name:0.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 36979.390625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 0.mlp.down_proj 1st kmeans time: 132.37139582633972
INFO - 0.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 0.mlp.down_proj proxy error before VPTQ: 3.37908460323888e-07, 3.972974809585139e-05, 0.00850517489016056
INFO - 0.mlp.down_proj 1st error time: 8.691988229751587
INFO - 0.mlp.down_proj proxy error after VPTQ: 1.5764405247864488e-07, 3.972974809585139e-05, 0.00396790960803628
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 14428.3330078125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 0.mlp.down_proj residual time: 1.0525047779083252
INFO - 0.mlp.down_proj 2ed gptq time: 9.496360063552856
INFO - 0.mlp.down_proj 2ed error time: 0.002386808395385742
INFO - 0.mlp.down_proj proxy error after residual VPTQ: 5.054142704352671e-08, 3.972974809585139e-05, 0.001272130524739623
INFO - ----Quantizing layer 1 ...---- 2026-02-23 08:16:12 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/1_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:16:13 1.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:1.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 87 iterations, error 1274.830078125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 1.self_attn.q_proj 1st kmeans time: 33.0862090587616
INFO - 1.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 1.self_attn.q_proj proxy error before VPTQ: 3.1764119512445177e-07, 0.00013008216046728194, 0.0024418504908680916
INFO - 1.self_attn.q_proj 1st error time: 2.266545295715332
INFO - 1.self_attn.q_proj proxy error after VPTQ: 1.6398787749949406e-07, 0.00013008216046728194, 0.0012606484815478325
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 69 iterations, error 537.1837158203125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 1.self_attn.q_proj residual time: 0.31831932067871094
INFO - 1.self_attn.q_proj 2ed gptq time: 2.563098669052124
INFO - 1.self_attn.q_proj 2ed error time: 0.0007650852203369141
INFO - 1.self_attn.q_proj proxy error after residual VPTQ: 5.6506689105617625e-08, 0.00013008216046728194, 0.0004343923064880073
INFO - load Hessian from /workspace/hessians/quip-format/hessians/1_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:16:53 1.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:1.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 23 iterations, error 190.72366333007812
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 1.self_attn.k_proj 1st kmeans time: 2.3623697757720947
INFO - 1.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 1.self_attn.k_proj proxy error before VPTQ: 4.3372985203404824e-08, 7.65465447329916e-05, 0.0005666223587468266
INFO - 1.self_attn.k_proj 1st error time: 1.5577735900878906
INFO - 1.self_attn.k_proj proxy error after VPTQ: 3.021943939529592e-08, 7.65465447329916e-05, 0.0003947851655539125
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 58 iterations, error 76.46443176269531
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 1.self_attn.k_proj residual time: 0.06088709831237793
INFO - 1.self_attn.k_proj 2ed gptq time: 1.955169677734375
INFO - 1.self_attn.k_proj 2ed error time: 0.00048041343688964844
INFO - 1.self_attn.k_proj proxy error after residual VPTQ: 9.58975920895e-09, 7.65465447329916e-05, 0.00012528010120149702
INFO - load Hessian from /workspace/hessians/quip-format/hessians/1_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:17:00 1.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:1.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 23 iterations, error 587.230224609375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 1.self_attn.v_proj 1st kmeans time: 2.3601572513580322
INFO - 1.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 1.self_attn.v_proj proxy error before VPTQ: 3.04419955909907e-08, 5.598084499069955e-06, 0.005437930580228567
INFO - 1.self_attn.v_proj 1st error time: 1.5003619194030762
INFO - 1.self_attn.v_proj proxy error after VPTQ: 1.3641147766918493e-08, 5.598084499069955e-06, 0.00243675266392529
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 81 iterations, error 214.52874755859375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 1.self_attn.v_proj residual time: 0.08240318298339844
INFO - 1.self_attn.v_proj 2ed gptq time: 2.0284600257873535
INFO - 1.self_attn.v_proj 2ed error time: 0.0004382133483886719
INFO - 1.self_attn.v_proj proxy error after residual VPTQ: 4.191956115562334e-09, 5.598084499069955e-06, 0.0007488197297789156
INFO - load Hessian from /workspace/hessians/quip-format/hessians/1_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:17:07 1.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:1.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 54 iterations, error 12323.15234375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 1.self_attn.o_proj 1st kmeans time: 20.84084677696228
INFO - 1.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 1.self_attn.o_proj proxy error before VPTQ: 2.1644948446919443e-06, 0.0010309251956641674, 0.002099565463140607
INFO - 1.self_attn.o_proj 1st error time: 2.275160551071167
INFO - 1.self_attn.o_proj proxy error after VPTQ: 1.294328512813081e-06, 0.0010309251956641674, 0.0012555018765851855
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 5325.05908203125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 1.self_attn.o_proj residual time: 0.31118345260620117
INFO - 1.self_attn.o_proj 2ed gptq time: 2.5588572025299072
INFO - 1.self_attn.o_proj 2ed error time: 0.0010538101196289062
INFO - 1.self_attn.o_proj proxy error after residual VPTQ: 4.452390953701979e-07, 0.0010309251956641674, 0.00043188303243368864
INFO - load Hessian from /workspace/hessians/quip-format/hessians/1_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:17:35 1.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:1.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 59173.5234375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 1.mlp.gate_proj 1st kmeans time: 132.48823046684265
INFO - 1.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 1.mlp.gate_proj proxy error before VPTQ: 1.4449969057750423e-05, 0.003676970023661852, 0.0039298576302826405
INFO - 1.mlp.gate_proj 1st error time: 4.78622841835022
INFO - 1.mlp.gate_proj proxy error after VPTQ: 8.14692975836806e-06, 0.003676970023661852, 0.0022156639024615288
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 24918.189453125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 1.mlp.gate_proj residual time: 1.014310598373413
INFO - 1.mlp.gate_proj 2ed gptq time: 5.0244140625
INFO - 1.mlp.gate_proj 2ed error time: 0.0016856193542480469
INFO - 1.mlp.gate_proj proxy error after residual VPTQ: 2.538979970267974e-06, 0.003676970023661852, 0.0006905087502673268
INFO - load Hessian from /workspace/hessians/quip-format/hessians/1_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:20:01 1.mlp.up_proj
INFO - enabling norm dim 0, layer_name:1.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 65606.4453125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 1.mlp.up_proj 1st kmeans time: 132.49833846092224
INFO - 1.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 1.mlp.up_proj proxy error before VPTQ: 1.1233906661800575e-05, 0.0012899402063339949, 0.008708858862519264
INFO - 1.mlp.up_proj 1st error time: 4.790868282318115
INFO - 1.mlp.up_proj proxy error after VPTQ: 5.969131052552257e-06, 0.0012899402063339949, 0.0046274480409920216
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 26942.775390625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 1.mlp.up_proj residual time: 1.0904645919799805
INFO - 1.mlp.up_proj 2ed gptq time: 5.022579908370972
INFO - 1.mlp.up_proj 2ed error time: 0.0012431144714355469
INFO - 1.mlp.up_proj proxy error after residual VPTQ: 1.8383761926088482e-06, 0.0012899402063339949, 0.0014251639368012547
INFO - load Hessian from /workspace/hessians/quip-format/hessians/1_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:22:29 1.mlp.down_proj
INFO - enabling norm dim 0, layer_name:1.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 996.4207763671875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 1.mlp.down_proj 1st kmeans time: 132.37474370002747
INFO - 1.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 1.mlp.down_proj proxy error before VPTQ: 8.385438121649713e-09, 8.644621266284958e-05, 9.700180089566857e-05
INFO - 1.mlp.down_proj 1st error time: 8.458587408065796
INFO - 1.mlp.down_proj proxy error after VPTQ: 8.082070124260099e-09, 8.644621266284958e-05, 9.349247557111084e-05
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 99 iterations, error 352.59912109375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 1.mlp.down_proj residual time: 1.0010621547698975
INFO - 1.mlp.down_proj 2ed gptq time: 9.490590333938599
INFO - 1.mlp.down_proj 2ed error time: 0.0022554397583007812
INFO - 1.mlp.down_proj proxy error after residual VPTQ: 2.6772741801295297e-09, 8.644621266284958e-05, 3.097040462307632e-05
INFO - ----Quantizing layer 2 ...---- 2026-02-23 08:25:05 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/2_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:25:06 2.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:2.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 54 iterations, error 5436.56640625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 2.self_attn.q_proj 1st kmeans time: 20.908185482025146
INFO - 2.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 2.self_attn.q_proj proxy error before VPTQ: 1.178298816739698e-06, 0.00013419587048701942, 0.00878044031560421
INFO - 2.self_attn.q_proj 1st error time: 2.2993316650390625
INFO - 2.self_attn.q_proj proxy error after VPTQ: 4.6326667302309943e-07, 0.00013419587048701942, 0.003452167846262455
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 85 iterations, error 2509.345947265625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 2.self_attn.q_proj residual time: 0.27207112312316895
INFO - 2.self_attn.q_proj 2ed gptq time: 2.558058500289917
INFO - 2.self_attn.q_proj 2ed error time: 0.0005719661712646484
INFO - 2.self_attn.q_proj proxy error after residual VPTQ: 1.5882048387538816e-07, 0.00013419587048701942, 0.0011834976030513644
INFO - load Hessian from /workspace/hessians/quip-format/hessians/2_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:25:33 2.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:2.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 27 iterations, error 991.6791381835938
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 2.self_attn.k_proj 1st kmeans time: 2.7363011837005615
INFO - 2.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 2.self_attn.k_proj proxy error before VPTQ: 2.2671453336897684e-07, 4.4117092329543084e-05, 0.0051389276050031185
INFO - 2.self_attn.k_proj 1st error time: 1.5458180904388428
INFO - 2.self_attn.k_proj proxy error after VPTQ: 1.2903645085771132e-07, 4.4117092329543084e-05, 0.002924863016232848
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 461.24041748046875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 2.self_attn.k_proj residual time: 0.08989644050598145
INFO - 2.self_attn.k_proj 2ed gptq time: 1.9648456573486328
INFO - 2.self_attn.k_proj 2ed error time: 0.0006196498870849609
INFO - 2.self_attn.k_proj proxy error after residual VPTQ: 4.3649258429923066e-08, 4.4117092329543084e-05, 0.000989395659416914
INFO - load Hessian from /workspace/hessians/quip-format/hessians/2_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:25:41 2.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:2.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 23 iterations, error 1615.7613525390625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 2.self_attn.v_proj 1st kmeans time: 2.367004632949829
INFO - 2.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 2.self_attn.v_proj proxy error before VPTQ: 7.233373366943852e-08, 4.467199687496759e-06, 0.016192186623811722
INFO - 2.self_attn.v_proj 1st error time: 1.5413403511047363
INFO - 2.self_attn.v_proj proxy error after VPTQ: 3.282585936403848e-08, 4.467199687496759e-06, 0.007348196115344763
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 75 iterations, error 744.4148559570312
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 2.self_attn.v_proj residual time: 0.07564234733581543
INFO - 2.self_attn.v_proj 2ed gptq time: 2.0491886138916016
INFO - 2.self_attn.v_proj 2ed error time: 0.00041174888610839844
INFO - 2.self_attn.v_proj proxy error after residual VPTQ: 1.064120613847308e-08, 4.467199687496759e-06, 0.0023820754140615463
INFO - load Hessian from /workspace/hessians/quip-format/hessians/2_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:25:49 2.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:2.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 59 iterations, error 15161.0595703125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 2.self_attn.o_proj 1st kmeans time: 22.715409517288208
INFO - 2.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 2.self_attn.o_proj proxy error before VPTQ: 1.5650658724553068e-06, 0.00014058996748644859, 0.011132130399346352
INFO - 2.self_attn.o_proj 1st error time: 2.287667751312256
INFO - 2.self_attn.o_proj proxy error after VPTQ: 9.915536338667152e-07, 0.00014058996748644859, 0.00705280527472496
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 6243.71484375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 2.self_attn.o_proj residual time: 0.30925416946411133
INFO - 2.self_attn.o_proj 2ed gptq time: 2.6312334537506104
INFO - 2.self_attn.o_proj 2ed error time: 0.0010051727294921875
INFO - 2.self_attn.o_proj proxy error after residual VPTQ: 3.236259260575025e-07, 0.00014058996748644859, 0.0023019134532660246
INFO - load Hessian from /workspace/hessians/quip-format/hessians/2_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:26:18 2.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:2.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 73170.59375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 2.mlp.gate_proj 1st kmeans time: 132.41144442558289
INFO - 2.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 2.mlp.gate_proj proxy error before VPTQ: 1.0103492058988195e-05, 0.0002298607723787427, 0.04395483434200287
INFO - 2.mlp.gate_proj 1st error time: 4.7974982261657715
INFO - 2.mlp.gate_proj proxy error after VPTQ: 6.656436198682059e-06, 0.0002298607723787427, 0.028958557173609734
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 27628.03125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 2.mlp.gate_proj residual time: 1.0141174793243408
INFO - 2.mlp.gate_proj 2ed gptq time: 5.063022613525391
INFO - 2.mlp.gate_proj 2ed error time: 0.0013146400451660156
INFO - 2.mlp.gate_proj proxy error after residual VPTQ: 2.0769844013557304e-06, 0.0002298607723787427, 0.009035836905241013
INFO - load Hessian from /workspace/hessians/quip-format/hessians/2_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:28:44 2.mlp.up_proj
INFO - enabling norm dim 0, layer_name:2.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 74616.1328125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 2.mlp.up_proj 1st kmeans time: 132.4623999595642
INFO - 2.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 2.mlp.up_proj proxy error before VPTQ: 8.337084182130639e-06, 0.0001508991845184937, 0.05524936690926552
INFO - 2.mlp.up_proj 1st error time: 4.863592863082886
INFO - 2.mlp.up_proj proxy error after VPTQ: 5.270862402539933e-06, 0.0001508991845184937, 0.03492969274520874
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 27889.19140625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 2.mlp.up_proj residual time: 1.0324647426605225
INFO - 2.mlp.up_proj 2ed gptq time: 5.075443267822266
INFO - 2.mlp.up_proj 2ed error time: 0.001354217529296875
INFO - 2.mlp.up_proj proxy error after residual VPTQ: 1.6353051250916906e-06, 0.0001508991845184937, 0.010837070643901825
INFO - load Hessian from /workspace/hessians/quip-format/hessians/2_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:31:12 2.mlp.down_proj
INFO - enabling norm dim 0, layer_name:2.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 80463.4453125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 2.mlp.down_proj 1st kmeans time: 132.4552984237671
INFO - 2.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 2.mlp.down_proj proxy error before VPTQ: 6.987883125475491e-07, 5.6217268138425425e-05, 0.01243013609200716
INFO - 2.mlp.down_proj 1st error time: 8.958659648895264
INFO - 2.mlp.down_proj proxy error after VPTQ: 4.659515013827331e-07, 5.6217268138425425e-05, 0.008288404904305935
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 29517.685546875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 2.mlp.down_proj residual time: 1.0800023078918457
INFO - 2.mlp.down_proj 2ed gptq time: 9.653671026229858
INFO - 2.mlp.down_proj 2ed error time: 0.002213716506958008
INFO - 2.mlp.down_proj proxy error after residual VPTQ: 1.5033556621801836e-07, 5.6217268138425425e-05, 0.0026741884648799896
INFO - ----Quantizing layer 3 ...---- 2026-02-23 08:33:49 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/3_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:33:49 3.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:3.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 62 iterations, error 5797.98046875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 3.self_attn.q_proj 1st kmeans time: 23.813682079315186
INFO - 3.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 3.self_attn.q_proj proxy error before VPTQ: 1.1571061122594983e-06, 0.00020506921282503754, 0.005642515141516924
INFO - 3.self_attn.q_proj 1st error time: 2.3387372493743896
INFO - 3.self_attn.q_proj proxy error after VPTQ: 5.008009225093701e-07, 0.00020506921282503754, 0.0024421068374067545
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 84 iterations, error 2554.467041015625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 3.self_attn.q_proj residual time: 0.3707554340362549
INFO - 3.self_attn.q_proj 2ed gptq time: 2.5974831581115723
INFO - 3.self_attn.q_proj 2ed error time: 0.0002703666687011719
INFO - 3.self_attn.q_proj proxy error after residual VPTQ: 1.6992547102745448e-07, 0.00020506921282503754, 0.0008286249940283597
INFO - load Hessian from /workspace/hessians/quip-format/hessians/3_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:34:20 3.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:3.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 28 iterations, error 984.5797729492188
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 3.self_attn.k_proj 1st kmeans time: 2.8283402919769287
INFO - 3.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 3.self_attn.k_proj proxy error before VPTQ: 2.365668194670434e-07, 6.766001752112061e-05, 0.0034964049700647593
INFO - 3.self_attn.k_proj 1st error time: 1.5660388469696045
INFO - 3.self_attn.k_proj proxy error after VPTQ: 1.4320664831757313e-07, 6.766001752112061e-05, 0.002116562332957983
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 435.8312683105469
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 3.self_attn.k_proj residual time: 0.09082984924316406
INFO - 3.self_attn.k_proj 2ed gptq time: 2.047436475753784
INFO - 3.self_attn.k_proj 2ed error time: 0.0003895759582519531
INFO - 3.self_attn.k_proj proxy error after residual VPTQ: 4.7076834874815177e-08, 6.766001752112061e-05, 0.0006957851583138108
INFO - load Hessian from /workspace/hessians/quip-format/hessians/3_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:34:28 3.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:3.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 1722.008056640625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 3.self_attn.v_proj 1st kmeans time: 2.2748095989227295
INFO - 3.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 3.self_attn.v_proj proxy error before VPTQ: 9.804369938137825e-08, 8.443829756288324e-06, 0.011611283756792545
INFO - 3.self_attn.v_proj 1st error time: 1.5452148914337158
INFO - 3.self_attn.v_proj proxy error after VPTQ: 4.604709147315589e-08, 8.443829756288324e-06, 0.005453342106193304
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 81 iterations, error 750.6530151367188
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 3.self_attn.v_proj residual time: 0.07686614990234375
INFO - 3.self_attn.v_proj 2ed gptq time: 2.0545132160186768
INFO - 3.self_attn.v_proj 2ed error time: 0.0004241466522216797
INFO - 3.self_attn.v_proj proxy error after residual VPTQ: 1.463917698174555e-08, 8.443829756288324e-06, 0.001733712968416512
INFO - load Hessian from /workspace/hessians/quip-format/hessians/3_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:34:36 3.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:3.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 64 iterations, error 18126.23046875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 3.self_attn.o_proj 1st kmeans time: 24.63776993751526
INFO - 3.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 3.self_attn.o_proj proxy error before VPTQ: 2.2872827685205266e-06, 0.0004646872985176742, 0.004922197666019201
INFO - 3.self_attn.o_proj 1st error time: 2.3152244091033936
INFO - 3.self_attn.o_proj proxy error after VPTQ: 1.5925628531476832e-06, 0.0004646872985176742, 0.0034271711483597755
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7173.5849609375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 3.self_attn.o_proj residual time: 0.31296873092651367
INFO - 3.self_attn.o_proj 2ed gptq time: 2.6554877758026123
INFO - 3.self_attn.o_proj 2ed error time: 0.001354217529296875
INFO - 3.self_attn.o_proj proxy error after residual VPTQ: 5.197754262553644e-07, 0.0004646872985176742, 0.0011185487965121865
INFO - load Hessian from /workspace/hessians/quip-format/hessians/3_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:35:07 3.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:3.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 72888.3984375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 3.mlp.gate_proj 1st kmeans time: 132.40184044837952
INFO - 3.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 3.mlp.gate_proj proxy error before VPTQ: 9.707629942568019e-06, 0.00020601661526598036, 0.04712061583995819
INFO - 3.mlp.gate_proj 1st error time: 4.749328136444092
INFO - 3.mlp.gate_proj proxy error after VPTQ: 6.754126843588892e-06, 0.00020601661526598036, 0.03278438001871109
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 26885.9296875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 3.mlp.gate_proj residual time: 1.0550193786621094
INFO - 3.mlp.gate_proj 2ed gptq time: 5.0546581745147705
INFO - 3.mlp.gate_proj 2ed error time: 0.0012099742889404297
INFO - 3.mlp.gate_proj proxy error after residual VPTQ: 2.123483682225924e-06, 0.00020601661526598036, 0.010307341814041138
INFO - load Hessian from /workspace/hessians/quip-format/hessians/3_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:37:33 3.mlp.up_proj
INFO - enabling norm dim 0, layer_name:3.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 74491.1640625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 3.mlp.up_proj 1st kmeans time: 132.40206599235535
INFO - 3.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 3.mlp.up_proj proxy error before VPTQ: 8.204089681385085e-06, 0.00013962676166556776, 0.05875728651881218
INFO - 3.mlp.up_proj 1st error time: 4.745235919952393
INFO - 3.mlp.up_proj proxy error after VPTQ: 5.6032663451333065e-06, 0.00013962676166556776, 0.04013031721115112
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 27214.34765625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 3.mlp.up_proj residual time: 1.0328154563903809
INFO - 3.mlp.up_proj 2ed gptq time: 5.049253702163696
INFO - 3.mlp.up_proj 2ed error time: 0.001110076904296875
INFO - 3.mlp.up_proj proxy error after residual VPTQ: 1.7581287465873174e-06, 0.00013962676166556776, 0.012591631151735783
INFO - load Hessian from /workspace/hessians/quip-format/hessians/3_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:40:00 3.mlp.down_proj
INFO - enabling norm dim 0, layer_name:3.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 86666.671875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 3.mlp.down_proj 1st kmeans time: 132.35540652275085
INFO - 3.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 3.mlp.down_proj proxy error before VPTQ: 6.638879312959034e-07, 2.024431523750536e-05, 0.0327937938272953
INFO - 3.mlp.down_proj 1st error time: 8.634475231170654
INFO - 3.mlp.down_proj proxy error after VPTQ: 4.899924874735007e-07, 2.024431523750536e-05, 0.0242039542645216
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 30869.12890625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 3.mlp.down_proj residual time: 1.0270068645477295
INFO - 3.mlp.down_proj 2ed gptq time: 9.613247156143188
INFO - 3.mlp.down_proj 2ed error time: 0.0025017261505126953
INFO - 3.mlp.down_proj proxy error after residual VPTQ: 1.5757737514832115e-07, 2.024431523750536e-05, 0.007783784065395594
INFO - ----Quantizing layer 4 ...---- 2026-02-23 08:42:36 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/4_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:42:37 4.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:4.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 60 iterations, error 6678.1826171875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 4.self_attn.q_proj 1st kmeans time: 23.064910173416138
INFO - 4.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 4.self_attn.q_proj proxy error before VPTQ: 1.3339936231204774e-06, 0.00010483874211786315, 0.012724243104457855
INFO - 4.self_attn.q_proj 1st error time: 2.3026671409606934
INFO - 4.self_attn.q_proj proxy error after VPTQ: 7.44808062336233e-07, 0.00010483874211786315, 0.007104320917278528
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 2913.55859375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 4.self_attn.q_proj residual time: 0.3189857006072998
INFO - 4.self_attn.q_proj 2ed gptq time: 2.590381383895874
INFO - 4.self_attn.q_proj 2ed error time: 0.00032210350036621094
INFO - 4.self_attn.q_proj proxy error after residual VPTQ: 2.574127222487732e-07, 0.00010483874211786315, 0.0024553206749260426
INFO - load Hessian from /workspace/hessians/quip-format/hessians/4_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:43:07 4.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:4.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 1312.4210205078125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 4.self_attn.k_proj 1st kmeans time: 2.26100492477417
INFO - 4.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 4.self_attn.k_proj proxy error before VPTQ: 2.884839318539889e-07, 4.1723418689798564e-05, 0.006914197001606226
INFO - 4.self_attn.k_proj 1st error time: 1.5221397876739502
INFO - 4.self_attn.k_proj proxy error after VPTQ: 1.9931626127345226e-07, 4.1723418689798564e-05, 0.0047770836390554905
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 586.9027709960938
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 4.self_attn.k_proj residual time: 0.0912008285522461
INFO - 4.self_attn.k_proj 2ed gptq time: 2.0593605041503906
INFO - 4.self_attn.k_proj 2ed error time: 0.00037670135498046875
INFO - 4.self_attn.k_proj proxy error after residual VPTQ: 6.710774869134184e-08, 4.1723418689798564e-05, 0.0016083952505141497
INFO - load Hessian from /workspace/hessians/quip-format/hessians/4_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:43:14 4.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:4.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 2063.072509765625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 4.self_attn.v_proj 1st kmeans time: 2.2600014209747314
INFO - 4.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 4.self_attn.v_proj proxy error before VPTQ: 1.1452539183665067e-07, 5.299185886542546e-06, 0.021611884236335754
INFO - 4.self_attn.v_proj 1st error time: 1.5426146984100342
INFO - 4.self_attn.v_proj proxy error after VPTQ: 6.73517419613745e-08, 5.299185886542546e-06, 0.012709828093647957
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 97 iterations, error 895.9276733398438
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 4.self_attn.v_proj residual time: 0.08892941474914551
INFO - 4.self_attn.v_proj 2ed gptq time: 2.049304246902466
INFO - 4.self_attn.v_proj 2ed error time: 0.0004048347473144531
INFO - 4.self_attn.v_proj proxy error after residual VPTQ: 2.1633105973251077e-08, 5.299185886542546e-06, 0.004082345403730869
INFO - load Hessian from /workspace/hessians/quip-format/hessians/4_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:43:22 4.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:4.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 74 iterations, error 17845.3203125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 4.self_attn.o_proj 1st kmeans time: 28.293111085891724
INFO - 4.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 4.self_attn.o_proj proxy error before VPTQ: 1.9268429696239764e-06, 0.0003161008935421705, 0.006095658056437969
INFO - 4.self_attn.o_proj 1st error time: 2.3210220336914062
INFO - 4.self_attn.o_proj proxy error after VPTQ: 1.2861802360930596e-06, 0.0003161008935421705, 0.004068891517817974
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7161.37109375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 4.self_attn.o_proj residual time: 0.3236370086669922
INFO - 4.self_attn.o_proj 2ed gptq time: 2.6086645126342773
INFO - 4.self_attn.o_proj 2ed error time: 0.0013196468353271484
INFO - 4.self_attn.o_proj proxy error after residual VPTQ: 4.167592351222993e-07, 0.0003161008935421705, 0.0013184373965486884
INFO - load Hessian from /workspace/hessians/quip-format/hessians/4_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:43:57 4.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:4.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 69551.3125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 4.mlp.gate_proj 1st kmeans time: 132.41009521484375
INFO - 4.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 4.mlp.gate_proj proxy error before VPTQ: 1.0075247701024637e-05, 0.00023761586635373533, 0.04240141063928604
INFO - 4.mlp.gate_proj 1st error time: 4.835468530654907
INFO - 4.mlp.gate_proj proxy error after VPTQ: 7.304595783352852e-06, 0.00023761586635373533, 0.030741196125745773
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 25642.712890625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 4.mlp.gate_proj residual time: 1.0715484619140625
INFO - 4.mlp.gate_proj 2ed gptq time: 5.049728870391846
INFO - 4.mlp.gate_proj 2ed error time: 0.001234292984008789
INFO - 4.mlp.gate_proj proxy error after residual VPTQ: 2.334640612389194e-06, 0.00023761586635373533, 0.009825272485613823
INFO - load Hessian from /workspace/hessians/quip-format/hessians/4_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:46:23 4.mlp.up_proj
INFO - enabling norm dim 0, layer_name:4.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 71393.859375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 4.mlp.up_proj 1st kmeans time: 132.41349983215332
INFO - 4.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 4.mlp.up_proj proxy error before VPTQ: 8.144320418068673e-06, 0.00014095182996243238, 0.057780876755714417
INFO - 4.mlp.up_proj 1st error time: 4.767737150192261
INFO - 4.mlp.up_proj proxy error after VPTQ: 5.796691766590811e-06, 0.00014095182996243238, 0.04112533852458
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 26065.05859375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 4.mlp.up_proj residual time: 1.114349603652954
INFO - 4.mlp.up_proj 2ed gptq time: 5.132198095321655
INFO - 4.mlp.up_proj 2ed error time: 0.0011508464813232422
INFO - 4.mlp.up_proj proxy error after residual VPTQ: 1.842482106440002e-06, 0.00014095182996243238, 0.013071714900434017
INFO - load Hessian from /workspace/hessians/quip-format/hessians/4_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:48:50 4.mlp.down_proj
INFO - enabling norm dim 0, layer_name:4.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 84887.515625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 4.mlp.down_proj 1st kmeans time: 132.42320013046265
INFO - 4.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 4.mlp.down_proj proxy error before VPTQ: 6.934282055226504e-07, 2.1484784156200476e-05, 0.03227531537413597
INFO - 4.mlp.down_proj 1st error time: 8.617223024368286
INFO - 4.mlp.down_proj proxy error after VPTQ: 5.089196406515839e-07, 2.1484784156200476e-05, 0.023687444627285004
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 30361.857421875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 4.mlp.down_proj residual time: 1.004713535308838
INFO - 4.mlp.down_proj 2ed gptq time: 9.62236762046814
INFO - 4.mlp.down_proj 2ed error time: 0.002254962921142578
INFO - 4.mlp.down_proj proxy error after residual VPTQ: 1.642922740074937e-07, 2.1484784156200476e-05, 0.007646912708878517
INFO - ----Quantizing layer 5 ...---- 2026-02-23 08:51:27 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/5_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:51:28 5.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:5.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 63 iterations, error 7785.8056640625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 5.self_attn.q_proj 1st kmeans time: 24.17437505722046
INFO - 5.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 5.self_attn.q_proj proxy error before VPTQ: 1.5357802567450562e-06, 0.00012591741688083857, 0.01219672616571188
INFO - 5.self_attn.q_proj 1st error time: 2.309755802154541
INFO - 5.self_attn.q_proj proxy error after VPTQ: 8.796039878689044e-07, 0.00012591741688083857, 0.006985562387853861
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3338.0302734375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 5.self_attn.q_proj residual time: 0.3108484745025635
INFO - 5.self_attn.q_proj 2ed gptq time: 2.599165678024292
INFO - 5.self_attn.q_proj 2ed error time: 0.0003151893615722656
INFO - 5.self_attn.q_proj proxy error after residual VPTQ: 2.993691055053205e-07, 0.00012591741688083857, 0.002377503551542759
INFO - load Hessian from /workspace/hessians/quip-format/hessians/5_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:51:59 5.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:5.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 23 iterations, error 1751.2650146484375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 5.self_attn.k_proj 1st kmeans time: 2.3550312519073486
INFO - 5.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 5.self_attn.k_proj proxy error before VPTQ: 3.4863012388086645e-07, 3.81335339625366e-05, 0.009142350405454636
INFO - 5.self_attn.k_proj 1st error time: 1.548957109451294
INFO - 5.self_attn.k_proj proxy error after VPTQ: 2.4311142965416366e-07, 3.81335339625366e-05, 0.0063752662390470505
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 760.0859375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 5.self_attn.k_proj residual time: 0.09172677993774414
INFO - 5.self_attn.k_proj 2ed gptq time: 2.0417842864990234
INFO - 5.self_attn.k_proj 2ed error time: 0.00042176246643066406
INFO - 5.self_attn.k_proj proxy error after residual VPTQ: 7.981896743558536e-08, 3.81335339625366e-05, 0.0020931437611579895
INFO - load Hessian from /workspace/hessians/quip-format/hessians/5_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:52:06 5.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:5.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 26 iterations, error 2599.594970703125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 5.self_attn.v_proj 1st kmeans time: 2.656126022338867
INFO - 5.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 5.self_attn.v_proj proxy error before VPTQ: 1.2760382617216237e-07, 6.558865607075859e-06, 0.019455166533589363
INFO - 5.self_attn.v_proj 1st error time: 1.543579339981079
INFO - 5.self_attn.v_proj proxy error after VPTQ: 7.49551816170424e-08, 6.558865607075859e-06, 0.011428071185946465
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1109.940185546875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 5.self_attn.v_proj residual time: 0.08960890769958496
INFO - 5.self_attn.v_proj 2ed gptq time: 2.0474131107330322
INFO - 5.self_attn.v_proj 2ed error time: 0.00037217140197753906
INFO - 5.self_attn.v_proj proxy error after residual VPTQ: 2.3685691630248584e-08, 6.558865607075859e-06, 0.003611248219385743
INFO - load Hessian from /workspace/hessians/quip-format/hessians/5_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:52:14 5.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:5.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 64 iterations, error 20200.560546875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 5.self_attn.o_proj 1st kmeans time: 24.637274503707886
INFO - 5.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 5.self_attn.o_proj proxy error before VPTQ: 1.6144776964210905e-06, 8.401586092077196e-05, 0.019216343760490417
INFO - 5.self_attn.o_proj 1st error time: 2.3542606830596924
INFO - 5.self_attn.o_proj proxy error after VPTQ: 1.0333669706596993e-06, 8.401586092077196e-05, 0.012299665249884129
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7761.56005859375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 5.self_attn.o_proj residual time: 0.31140732765197754
INFO - 5.self_attn.o_proj 2ed gptq time: 2.5959973335266113
INFO - 5.self_attn.o_proj 2ed error time: 0.001102447509765625
INFO - 5.self_attn.o_proj proxy error after residual VPTQ: 3.2775574254628737e-07, 8.401586092077196e-05, 0.003901117481291294
INFO - load Hessian from /workspace/hessians/quip-format/hessians/5_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:52:45 5.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:5.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 69622.640625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 5.mlp.gate_proj 1st kmeans time: 132.50213027000427
INFO - 5.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 5.mlp.gate_proj proxy error before VPTQ: 1.0231461601506453e-05, 0.00020316403242759407, 0.05036059394478798
INFO - 5.mlp.gate_proj 1st error time: 4.820861101150513
INFO - 5.mlp.gate_proj proxy error after VPTQ: 7.50948811401031e-06, 0.00020316403242759407, 0.036962684243917465
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 25606.595703125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 5.mlp.gate_proj residual time: 1.0733590126037598
INFO - 5.mlp.gate_proj 2ed gptq time: 5.093791484832764
INFO - 5.mlp.gate_proj 2ed error time: 0.0011370182037353516
INFO - 5.mlp.gate_proj proxy error after residual VPTQ: 2.4013393158384133e-06, 0.00020316403242759407, 0.011819707229733467
INFO - load Hessian from /workspace/hessians/quip-format/hessians/5_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:55:11 5.mlp.up_proj
INFO - enabling norm dim 0, layer_name:5.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 71223.40625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 5.mlp.up_proj 1st kmeans time: 132.39888048171997
INFO - 5.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 5.mlp.up_proj proxy error before VPTQ: 7.977571840456221e-06, 0.00013599048543255776, 0.058662720024585724
INFO - 5.mlp.up_proj 1st error time: 4.767603158950806
INFO - 5.mlp.up_proj proxy error after VPTQ: 5.734817932534497e-06, 0.00013599048543255776, 0.042170729488134384
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 99 iterations, error 25985.41015625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 5.mlp.up_proj residual time: 1.220475435256958
INFO - 5.mlp.up_proj 2ed gptq time: 5.122341156005859
INFO - 5.mlp.up_proj 2ed error time: 0.0010800361633300781
INFO - 5.mlp.up_proj proxy error after residual VPTQ: 1.8293183075002162e-06, 0.00013599048543255776, 0.01345181092619896
INFO - load Hessian from /workspace/hessians/quip-format/hessians/5_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 08:57:39 5.mlp.down_proj
INFO - enabling norm dim 0, layer_name:5.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 86157.2890625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 5.mlp.down_proj 1st kmeans time: 132.4708023071289
INFO - 5.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 5.mlp.down_proj proxy error before VPTQ: 6.68341101572878e-07, 2.0735531506943516e-05, 0.032231684774160385
INFO - 5.mlp.down_proj 1st error time: 8.62990927696228
INFO - 5.mlp.down_proj proxy error after VPTQ: 5.076909133094887e-07, 2.0735531506943516e-05, 0.02448410354554653
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 30538.2578125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 5.mlp.down_proj residual time: 1.041076421737671
INFO - 5.mlp.down_proj 2ed gptq time: 9.609449625015259
INFO - 5.mlp.down_proj 2ed error time: 0.002092599868774414
INFO - 5.mlp.down_proj proxy error after residual VPTQ: 1.639654811924629e-07, 2.0735531506943516e-05, 0.007907464168965816
INFO - ----Quantizing layer 6 ...---- 2026-02-23 09:00:15 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/6_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:00:15 6.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:6.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 63 iterations, error 7523.38623046875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 6.self_attn.q_proj 1st kmeans time: 24.181918382644653
INFO - 6.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 6.self_attn.q_proj proxy error before VPTQ: 1.4215130477168714e-06, 0.00012325581337790936, 0.011533030308783054
INFO - 6.self_attn.q_proj 1st error time: 2.3773601055145264
INFO - 6.self_attn.q_proj proxy error after VPTQ: 8.021293069759849e-07, 0.00012325581337790936, 0.006507841404527426
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3225.826171875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 6.self_attn.q_proj residual time: 0.31668758392333984
INFO - 6.self_attn.q_proj 2ed gptq time: 2.595366954803467
INFO - 6.self_attn.q_proj 2ed error time: 0.0002911090850830078
INFO - 6.self_attn.q_proj proxy error after residual VPTQ: 2.745798610703787e-07, 0.00012325581337790936, 0.002227723365649581
INFO - load Hessian from /workspace/hessians/quip-format/hessians/6_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:00:47 6.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:6.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 30 iterations, error 1495.457763671875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 6.self_attn.k_proj 1st kmeans time: 3.013547658920288
INFO - 6.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 6.self_attn.k_proj proxy error before VPTQ: 3.3545774158483255e-07, 4.505653487285599e-05, 0.007445262745022774
INFO - 6.self_attn.k_proj 1st error time: 1.5378375053405762
INFO - 6.self_attn.k_proj proxy error after VPTQ: 2.3707931973149243e-07, 4.505653487285599e-05, 0.005261818878352642
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 660.7884521484375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 6.self_attn.k_proj residual time: 0.09090995788574219
INFO - 6.self_attn.k_proj 2ed gptq time: 2.046696662902832
INFO - 6.self_attn.k_proj 2ed error time: 0.0003459453582763672
INFO - 6.self_attn.k_proj proxy error after residual VPTQ: 7.921369160612812e-08, 4.505653487285599e-05, 0.0017580954590812325
INFO - load Hessian from /workspace/hessians/quip-format/hessians/6_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:00:55 6.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:6.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 29 iterations, error 2369.47021484375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 6.self_attn.v_proj 1st kmeans time: 2.924139976501465
INFO - 6.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 6.self_attn.v_proj proxy error before VPTQ: 1.163336591503139e-07, 5.34519813299994e-06, 0.02176414430141449
INFO - 6.self_attn.v_proj 1st error time: 1.5600624084472656
INFO - 6.self_attn.v_proj proxy error after VPTQ: 7.063816553909419e-08, 5.34519813299994e-06, 0.013215256854891777
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1011.5257568359375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 6.self_attn.v_proj residual time: 0.10580945014953613
INFO - 6.self_attn.v_proj 2ed gptq time: 2.0503365993499756
INFO - 6.self_attn.v_proj 2ed error time: 0.0003902912139892578
INFO - 6.self_attn.v_proj proxy error after residual VPTQ: 2.256485842622169e-08, 5.34519813299994e-06, 0.0042215194553136826
INFO - load Hessian from /workspace/hessians/quip-format/hessians/6_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:01:03 6.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:6.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 85 iterations, error 17777.875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 6.self_attn.o_proj 1st kmeans time: 32.42554759979248
INFO - 6.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 6.self_attn.o_proj proxy error before VPTQ: 1.6883780062926235e-06, 8.733928552828729e-05, 0.019331255927681923
INFO - 6.self_attn.o_proj 1st error time: 2.348013401031494
INFO - 6.self_attn.o_proj proxy error after VPTQ: 1.0703536190703744e-06, 8.733928552828729e-05, 0.01225512195378542
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 6963.8798828125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 6.self_attn.o_proj residual time: 0.31485724449157715
INFO - 6.self_attn.o_proj 2ed gptq time: 2.6005115509033203
INFO - 6.self_attn.o_proj 2ed error time: 0.001216888427734375
INFO - 6.self_attn.o_proj proxy error after residual VPTQ: 3.486713922029594e-07, 8.733928552828729e-05, 0.003992148209363222
INFO - load Hessian from /workspace/hessians/quip-format/hessians/6_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:01:42 6.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:6.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 66755.8515625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 6.mlp.gate_proj 1st kmeans time: 132.40297865867615
INFO - 6.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 6.mlp.gate_proj proxy error before VPTQ: 9.636491085984744e-06, 0.00020396191393956542, 0.04724652320146561
INFO - 6.mlp.gate_proj 1st error time: 4.756271123886108
INFO - 6.mlp.gate_proj proxy error after VPTQ: 7.06275568518322e-06, 0.00020396191393956542, 0.03462781757116318
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 24598.251953125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 6.mlp.gate_proj residual time: 1.0405683517456055
INFO - 6.mlp.gate_proj 2ed gptq time: 5.054155588150024
INFO - 6.mlp.gate_proj 2ed error time: 0.0010738372802734375
INFO - 6.mlp.gate_proj proxy error after residual VPTQ: 2.273114660056308e-06, 0.00020396191393956542, 0.011144799180328846
INFO - load Hessian from /workspace/hessians/quip-format/hessians/6_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:04:08 6.mlp.up_proj
INFO - enabling norm dim 0, layer_name:6.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 68225.734375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 6.mlp.up_proj 1st kmeans time: 132.4085385799408
INFO - 6.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 6.mlp.up_proj proxy error before VPTQ: 7.856708180042915e-06, 0.00012772015179507434, 0.061515022069215775
INFO - 6.mlp.up_proj 1st error time: 4.821458578109741
INFO - 6.mlp.up_proj proxy error after VPTQ: 5.71498731005704e-06, 0.00012772015179507434, 0.044746167957782745
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 24945.5859375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 6.mlp.up_proj residual time: 1.0641541481018066
INFO - 6.mlp.up_proj 2ed gptq time: 5.135112762451172
INFO - 6.mlp.up_proj 2ed error time: 0.0012235641479492188
INFO - 6.mlp.up_proj proxy error after residual VPTQ: 1.8347034256294137e-06, 0.00012772015179507434, 0.014365026727318764
INFO - load Hessian from /workspace/hessians/quip-format/hessians/6_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:06:36 6.mlp.down_proj
INFO - enabling norm dim 0, layer_name:6.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 85017.59375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 6.mlp.down_proj 1st kmeans time: 132.463604927063
INFO - 6.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 6.mlp.down_proj proxy error before VPTQ: 6.849118108220864e-07, 1.7965285223908722e-05, 0.03812418133020401
INFO - 6.mlp.down_proj 1st error time: 8.642677545547485
INFO - 6.mlp.down_proj proxy error after VPTQ: 5.27146482909302e-07, 1.7965285223908722e-05, 0.02934250608086586
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 30195.693359375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 6.mlp.down_proj residual time: 1.0293312072753906
INFO - 6.mlp.down_proj 2ed gptq time: 9.653807163238525
INFO - 6.mlp.down_proj 2ed error time: 0.0026922225952148438
INFO - 6.mlp.down_proj proxy error after residual VPTQ: 1.7112442662892136e-07, 1.7965285223908722e-05, 0.009525283239781857
INFO - ----Quantizing layer 7 ...---- 2026-02-23 09:09:12 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/7_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:09:13 7.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:7.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 61 iterations, error 7814.7841796875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 7.self_attn.q_proj 1st kmeans time: 23.44423007965088
INFO - 7.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 7.self_attn.q_proj proxy error before VPTQ: 1.5238769037750899e-06, 0.0001855981390690431, 0.008210626430809498
INFO - 7.self_attn.q_proj 1st error time: 2.306312084197998
INFO - 7.self_attn.q_proj proxy error after VPTQ: 8.256777164206142e-07, 0.0001855981390690431, 0.004448739346116781
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3397.135986328125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 7.self_attn.q_proj residual time: 0.3114488124847412
INFO - 7.self_attn.q_proj 2ed gptq time: 2.605501413345337
INFO - 7.self_attn.q_proj 2ed error time: 0.0003364086151123047
INFO - 7.self_attn.q_proj proxy error after residual VPTQ: 2.8690379849649617e-07, 0.0001855981390690431, 0.0015458334237337112
INFO - load Hessian from /workspace/hessians/quip-format/hessians/7_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:09:43 7.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:7.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 26 iterations, error 1454.3472900390625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 7.self_attn.k_proj 1st kmeans time: 2.637143135070801
INFO - 7.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 7.self_attn.k_proj proxy error before VPTQ: 3.3310027447441826e-07, 3.748614471987821e-05, 0.008885957300662994
INFO - 7.self_attn.k_proj 1st error time: 1.5305383205413818
INFO - 7.self_attn.k_proj proxy error after VPTQ: 2.4374577378694084e-07, 3.748614471987821e-05, 0.006502289790660143
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 659.6876831054688
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 7.self_attn.k_proj residual time: 0.09101986885070801
INFO - 7.self_attn.k_proj 2ed gptq time: 2.036165475845337
INFO - 7.self_attn.k_proj 2ed error time: 0.00037026405334472656
INFO - 7.self_attn.k_proj proxy error after residual VPTQ: 8.225698877595278e-08, 3.748614471987821e-05, 0.0021943305619060993
INFO - load Hessian from /workspace/hessians/quip-format/hessians/7_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:09:51 7.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:7.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 25 iterations, error 2144.88916015625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 7.self_attn.v_proj 1st kmeans time: 2.555532932281494
INFO - 7.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 7.self_attn.v_proj proxy error before VPTQ: 1.1391165344321053e-07, 4.806826837011613e-06, 0.02369789034128189
INFO - 7.self_attn.v_proj 1st error time: 1.5426232814788818
INFO - 7.self_attn.v_proj proxy error after VPTQ: 7.426412196309684e-08, 4.806826837011613e-06, 0.015449718572199345
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 930.201416015625
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 7.self_attn.v_proj residual time: 0.09138727188110352
INFO - 7.self_attn.v_proj 2ed gptq time: 2.0522828102111816
INFO - 7.self_attn.v_proj 2ed error time: 0.00035953521728515625
INFO - 7.self_attn.v_proj proxy error after residual VPTQ: 2.3883750088771194e-08, 4.806826837011613e-06, 0.004968714434653521
INFO - load Hessian from /workspace/hessians/quip-format/hessians/7_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:09:59 7.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:7.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 71 iterations, error 18476.3203125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 7.self_attn.o_proj 1st kmeans time: 27.165563106536865
INFO - 7.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 7.self_attn.o_proj proxy error before VPTQ: 1.5367536434496287e-06, 6.0655776906060055e-05, 0.025335650891065598
INFO - 7.self_attn.o_proj 1st error time: 2.308464288711548
INFO - 7.self_attn.o_proj proxy error after VPTQ: 9.481242386755184e-07, 6.0655776906060055e-05, 0.01563122682273388
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7163.423828125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 7.self_attn.o_proj residual time: 0.31465935707092285
INFO - 7.self_attn.o_proj 2ed gptq time: 2.5964534282684326
INFO - 7.self_attn.o_proj 2ed error time: 0.0010788440704345703
INFO - 7.self_attn.o_proj proxy error after residual VPTQ: 3.064770908167702e-07, 6.0655776906060055e-05, 0.005052727181464434
INFO - load Hessian from /workspace/hessians/quip-format/hessians/7_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:10:33 7.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:7.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 67412.625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 7.mlp.gate_proj 1st kmeans time: 132.39823031425476
INFO - 7.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 7.mlp.gate_proj proxy error before VPTQ: 9.938548828358762e-06, 0.0001923601666931063, 0.05166635662317276
INFO - 7.mlp.gate_proj 1st error time: 4.749197483062744
INFO - 7.mlp.gate_proj proxy error after VPTQ: 7.468994226655923e-06, 0.0001923601666931063, 0.03882817551493645
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 24742.388671875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 7.mlp.gate_proj residual time: 1.0518584251403809
INFO - 7.mlp.gate_proj 2ed gptq time: 5.127107858657837
INFO - 7.mlp.gate_proj 2ed error time: 0.0011737346649169922
INFO - 7.mlp.gate_proj proxy error after residual VPTQ: 2.4014470909605734e-06, 0.0001923601666931063, 0.012484118342399597
INFO - load Hessian from /workspace/hessians/quip-format/hessians/7_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:12:59 7.mlp.up_proj
INFO - enabling norm dim 0, layer_name:7.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 69052.59375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 7.mlp.up_proj 1st kmeans time: 132.41176533699036
INFO - 7.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 7.mlp.up_proj proxy error before VPTQ: 8.18792250356637e-06, 0.00014271728286985308, 0.05737162381410599
INFO - 7.mlp.up_proj 1st error time: 4.761176824569702
INFO - 7.mlp.up_proj proxy error after VPTQ: 6.084928372729337e-06, 0.00014271728286985308, 0.042636241763830185
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 89 iterations, error 25123.75390625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 7.mlp.up_proj residual time: 0.9325611591339111
INFO - 7.mlp.up_proj 2ed gptq time: 5.05283784866333
INFO - 7.mlp.up_proj 2ed error time: 0.0012812614440917969
INFO - 7.mlp.up_proj proxy error after residual VPTQ: 1.954384742930415e-06, 0.00014271728286985308, 0.013694100081920624
INFO - load Hessian from /workspace/hessians/quip-format/hessians/7_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:15:26 7.mlp.down_proj
INFO - enabling norm dim 0, layer_name:7.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 85993.34375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 7.mlp.down_proj 1st kmeans time: 132.57482957839966
INFO - 7.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 7.mlp.down_proj proxy error before VPTQ: 6.647223926847801e-07, 1.575286478328053e-05, 0.04219692200422287
INFO - 7.mlp.down_proj 1st error time: 8.606746912002563
INFO - 7.mlp.down_proj proxy error after VPTQ: 5.09900928591378e-07, 1.575286478328053e-05, 0.032368775457143784
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 30333.0234375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 7.mlp.down_proj residual time: 1.0025339126586914
INFO - 7.mlp.down_proj 2ed gptq time: 9.623483180999756
INFO - 7.mlp.down_proj 2ed error time: 0.002373218536376953
INFO - 7.mlp.down_proj proxy error after residual VPTQ: 1.646522775899939e-07, 1.575286478328053e-05, 0.0104522118344903
INFO - ----Quantizing layer 8 ...---- 2026-02-23 09:18:02 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/8_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:18:03 8.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:8.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 65 iterations, error 8810.962890625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 8.self_attn.q_proj 1st kmeans time: 24.93098759651184
INFO - 8.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 8.self_attn.q_proj proxy error before VPTQ: 1.520144678579527e-06, 0.00012858983245678246, 0.011821655556559563
INFO - 8.self_attn.q_proj 1st error time: 2.32967472076416
INFO - 8.self_attn.q_proj proxy error after VPTQ: 8.96142353212781e-07, 0.00012858983245678246, 0.006968998350203037
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3677.968994140625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 8.self_attn.q_proj residual time: 0.32126474380493164
INFO - 8.self_attn.q_proj 2ed gptq time: 2.598886251449585
INFO - 8.self_attn.q_proj 2ed error time: 0.0003476142883300781
INFO - 8.self_attn.q_proj proxy error after residual VPTQ: 3.027060131444159e-07, 0.00012858983245678246, 0.0023540430702269077
INFO - load Hessian from /workspace/hessians/quip-format/hessians/8_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:18:35 8.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:8.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 33 iterations, error 1723.49853515625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 8.self_attn.k_proj 1st kmeans time: 3.308659553527832
INFO - 8.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 8.self_attn.k_proj proxy error before VPTQ: 3.804811683494336e-07, 4.407561937114224e-05, 0.008632463403046131
INFO - 8.self_attn.k_proj 1st error time: 1.574697732925415
INFO - 8.self_attn.k_proj proxy error after VPTQ: 2.7486242970553576e-07, 4.407561937114224e-05, 0.006236155983060598
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 765.2239990234375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 8.self_attn.k_proj residual time: 0.09597516059875488
INFO - 8.self_attn.k_proj 2ed gptq time: 2.0296568870544434
INFO - 8.self_attn.k_proj 2ed error time: 0.00043082237243652344
INFO - 8.self_attn.k_proj proxy error after residual VPTQ: 9.155905189572877e-08, 4.407561937114224e-05, 0.0020773173309862614
INFO - load Hessian from /workspace/hessians/quip-format/hessians/8_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:18:43 8.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:8.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 23 iterations, error 2195.902587890625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 8.self_attn.v_proj 1st kmeans time: 2.3759849071502686
INFO - 8.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 8.self_attn.v_proj proxy error before VPTQ: 1.4276824344960914e-07, 5.7581760302127805e-06, 0.02479400485754013
INFO - 8.self_attn.v_proj 1st error time: 1.5487592220306396
INFO - 8.self_attn.v_proj proxy error after VPTQ: 9.816601220791199e-08, 5.7581760302127805e-06, 0.01704810932278633
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 950.722412109375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 8.self_attn.v_proj residual time: 0.09487462043762207
INFO - 8.self_attn.v_proj 2ed gptq time: 2.0488274097442627
INFO - 8.self_attn.v_proj 2ed error time: 0.0003466606140136719
INFO - 8.self_attn.v_proj proxy error after residual VPTQ: 3.218501376522909e-08, 5.7581760302127805e-06, 0.005589446052908897
INFO - load Hessian from /workspace/hessians/quip-format/hessians/8_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:18:51 8.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:8.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 79 iterations, error 18051.76171875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 8.self_attn.o_proj 1st kmeans time: 30.182859420776367
INFO - 8.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 8.self_attn.o_proj proxy error before VPTQ: 1.600959649294964e-06, 4.798247755388729e-05, 0.033365506678819656
INFO - 8.self_attn.o_proj 1st error time: 2.314793348312378
INFO - 8.self_attn.o_proj proxy error after VPTQ: 9.337356914329575e-07, 4.798247755388729e-05, 0.019459931179881096
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7048.8564453125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 8.self_attn.o_proj residual time: 0.3137974739074707
INFO - 8.self_attn.o_proj 2ed gptq time: 2.598954677581787
INFO - 8.self_attn.o_proj 2ed error time: 0.0011773109436035156
INFO - 8.self_attn.o_proj proxy error after residual VPTQ: 3.000475032877148e-07, 4.798247755388729e-05, 0.006253272294998169
INFO - load Hessian from /workspace/hessians/quip-format/hessians/8_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:19:28 8.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:8.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 66473.8984375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 8.mlp.gate_proj 1st kmeans time: 132.40039253234863
INFO - 8.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 8.mlp.gate_proj proxy error before VPTQ: 9.677731213741936e-06, 0.00021353407646529377, 0.045321717858314514
INFO - 8.mlp.gate_proj 1st error time: 4.7551984786987305
INFO - 8.mlp.gate_proj proxy error after VPTQ: 7.2011516749626026e-06, 0.00021353407646529377, 0.03372366353869438
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 99 iterations, error 24308.9921875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 8.mlp.gate_proj residual time: 1.0450677871704102
INFO - 8.mlp.gate_proj 2ed gptq time: 5.131145000457764
INFO - 8.mlp.gate_proj 2ed error time: 0.0011186599731445312
INFO - 8.mlp.gate_proj proxy error after residual VPTQ: 2.320459998372826e-06, 0.00021353407646529377, 0.010866930708289146
INFO - load Hessian from /workspace/hessians/quip-format/hessians/8_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:21:54 8.mlp.up_proj
INFO - enabling norm dim 0, layer_name:8.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 68123.4296875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 8.mlp.up_proj 1st kmeans time: 132.4171805381775
INFO - 8.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 8.mlp.up_proj proxy error before VPTQ: 7.935941539471969e-06, 0.00014134707453195006, 0.0561450719833374
INFO - 8.mlp.up_proj 1st error time: 4.774647951126099
INFO - 8.mlp.up_proj proxy error after VPTQ: 5.8482073654886335e-06, 0.00014134707453195006, 0.041374802589416504
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 24748.701171875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 8.mlp.up_proj residual time: 1.0163755416870117
INFO - 8.mlp.up_proj 2ed gptq time: 5.128051042556763
INFO - 8.mlp.up_proj 2ed error time: 0.0010828971862792969
INFO - 8.mlp.up_proj proxy error after residual VPTQ: 1.8800810721586458e-06, 0.00014134707453195006, 0.013301167637109756
INFO - load Hessian from /workspace/hessians/quip-format/hessians/8_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:24:21 8.mlp.down_proj
INFO - enabling norm dim 0, layer_name:8.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 87618.328125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 8.mlp.down_proj 1st kmeans time: 132.37019872665405
INFO - 8.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 8.mlp.down_proj proxy error before VPTQ: 6.873964366604923e-07, 1.6342110029654577e-05, 0.04206289350986481
INFO - 8.mlp.down_proj 1st error time: 8.870804786682129
INFO - 8.mlp.down_proj proxy error after VPTQ: 5.352304128791729e-07, 1.6342110029654577e-05, 0.03275160863995552
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 30755.931640625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 8.mlp.down_proj residual time: 1.0768563747406006
INFO - 8.mlp.down_proj 2ed gptq time: 9.58186388015747
INFO - 8.mlp.down_proj 2ed error time: 0.0024895668029785156
INFO - 8.mlp.down_proj proxy error after residual VPTQ: 1.7236253313512861e-07, 1.6342110029654577e-05, 0.01054714061319828
INFO - ----Quantizing layer 9 ...---- 2026-02-23 09:26:58 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/9_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:26:59 9.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:9.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 86 iterations, error 8460.9853515625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 9.self_attn.q_proj 1st kmeans time: 32.71372056007385
INFO - 9.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 9.self_attn.q_proj proxy error before VPTQ: 1.661784835960134e-06, 9.813340147957206e-05, 0.01693393662571907
INFO - 9.self_attn.q_proj 1st error time: 2.3204948902130127
INFO - 9.self_attn.q_proj proxy error after VPTQ: 1.0994532431141124e-06, 9.813340147957206e-05, 0.011203659698367119
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3594.005126953125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 9.self_attn.q_proj residual time: 0.31294989585876465
INFO - 9.self_attn.q_proj 2ed gptq time: 2.5929367542266846
INFO - 9.self_attn.q_proj 2ed error time: 0.00036215782165527344
INFO - 9.self_attn.q_proj proxy error after residual VPTQ: 3.795403245021589e-07, 9.813340147957206e-05, 0.0038675956893712282
INFO - load Hessian from /workspace/hessians/quip-format/hessians/9_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:27:38 9.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:9.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 26 iterations, error 1990.36083984375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 9.self_attn.k_proj 1st kmeans time: 2.6419801712036133
INFO - 9.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 9.self_attn.k_proj proxy error before VPTQ: 4.216503839415964e-07, 2.797174602164887e-05, 0.015074153430759907
INFO - 9.self_attn.k_proj 1st error time: 1.5512137413024902
INFO - 9.self_attn.k_proj proxy error after VPTQ: 3.178849112828175e-07, 2.797174602164887e-05, 0.011364500038325787
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 871.6823120117188
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 9.self_attn.k_proj residual time: 0.10879778861999512
INFO - 9.self_attn.k_proj 2ed gptq time: 2.055812358856201
INFO - 9.self_attn.k_proj 2ed error time: 0.0004525184631347656
INFO - 9.self_attn.k_proj proxy error after residual VPTQ: 1.0653523219161798e-07, 2.797174602164887e-05, 0.0038086730055510998
INFO - load Hessian from /workspace/hessians/quip-format/hessians/9_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:27:46 9.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:9.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 23 iterations, error 2745.1015625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 9.self_attn.v_proj 1st kmeans time: 2.4408347606658936
INFO - 9.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 9.self_attn.v_proj proxy error before VPTQ: 1.3217454863934108e-07, 5.250491540209623e-06, 0.025173747912049294
INFO - 9.self_attn.v_proj 1st error time: 1.5513484477996826
INFO - 9.self_attn.v_proj proxy error after VPTQ: 9.457337313278913e-08, 5.250491540209623e-06, 0.018012288957834244
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 100 iterations, error 1152.312255859375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 9.self_attn.v_proj residual time: 0.09138679504394531
INFO - 9.self_attn.v_proj 2ed gptq time: 2.0469603538513184
INFO - 9.self_attn.v_proj 2ed error time: 0.0004780292510986328
INFO - 9.self_attn.v_proj proxy error after residual VPTQ: 3.0211580792638415e-08, 5.250491540209623e-06, 0.005754048004746437
INFO - load Hessian from /workspace/hessians/quip-format/hessians/9_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:27:54 9.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:9.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 98 iterations, error 19894.767578125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 9.self_attn.o_proj 1st kmeans time: 37.17062187194824
INFO - 9.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 9.self_attn.o_proj proxy error before VPTQ: 1.5081844821906998e-06, 6.561295595020056e-05, 0.022986076772212982
INFO - 9.self_attn.o_proj 1st error time: 2.311878204345703
INFO - 9.self_attn.o_proj proxy error after VPTQ: 8.362104608750087e-07, 6.561295595020056e-05, 0.012744593434035778
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7568.8095703125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 9.self_attn.o_proj residual time: 0.3121664524078369
INFO - 9.self_attn.o_proj 2ed gptq time: 2.669680595397949
INFO - 9.self_attn.o_proj 2ed error time: 0.001168966293334961
INFO - 9.self_attn.o_proj proxy error after residual VPTQ: 2.663173290784471e-07, 6.561295595020056e-05, 0.004058913793414831
INFO - load Hessian from /workspace/hessians/quip-format/hessians/9_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:28:38 9.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:9.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 65647.2890625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 9.mlp.gate_proj 1st kmeans time: 132.40232253074646
INFO - 9.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 9.mlp.gate_proj proxy error before VPTQ: 9.632900400902145e-06, 0.00018794857896864414, 0.05125284940004349
INFO - 9.mlp.gate_proj 1st error time: 4.757849454879761
INFO - 9.mlp.gate_proj proxy error after VPTQ: 7.157336312957341e-06, 0.00018794857896864414, 0.03808135539293289
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 24084.626953125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 9.mlp.gate_proj residual time: 1.0402777194976807
INFO - 9.mlp.gate_proj 2ed gptq time: 5.106997966766357
INFO - 9.mlp.gate_proj 2ed error time: 0.0012030601501464844
INFO - 9.mlp.gate_proj proxy error after residual VPTQ: 2.307377371835173e-06, 0.00018794857896864414, 0.01227664202451706
INFO - load Hessian from /workspace/hessians/quip-format/hessians/9_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:31:03 9.mlp.up_proj
INFO - enabling norm dim 0, layer_name:9.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 67492.9296875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 9.mlp.up_proj 1st kmeans time: 132.40681099891663
INFO - 9.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 9.mlp.up_proj proxy error before VPTQ: 8.041097316890955e-06, 0.00016662782581988722, 0.04825783148407936
INFO - 9.mlp.up_proj 1st error time: 4.8518970012664795
INFO - 9.mlp.up_proj proxy error after VPTQ: 5.9280532695993315e-06, 0.00016662782581988722, 0.03557661175727844
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 24605.044921875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 9.mlp.up_proj residual time: 1.0384435653686523
INFO - 9.mlp.up_proj 2ed gptq time: 5.147325277328491
INFO - 9.mlp.up_proj 2ed error time: 0.0012459754943847656
INFO - 9.mlp.up_proj proxy error after residual VPTQ: 1.906011561914056e-06, 0.00016662782581988722, 0.011438734829425812
INFO - load Hessian from /workspace/hessians/quip-format/hessians/9_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:33:31 9.mlp.down_proj
INFO - enabling norm dim 0, layer_name:9.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 86945.671875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 9.mlp.down_proj 1st kmeans time: 132.46561765670776
INFO - 9.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 9.mlp.down_proj proxy error before VPTQ: 6.7404982928565e-07, 1.7052871044143103e-05, 0.039527058601379395
INFO - 9.mlp.down_proj 1st error time: 8.643865823745728
INFO - 9.mlp.down_proj proxy error after VPTQ: 5.254135544419114e-07, 1.7052871044143103e-05, 0.030810855329036713
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 30550.48828125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 9.mlp.down_proj residual time: 1.012115716934204
INFO - 9.mlp.down_proj 2ed gptq time: 9.646767139434814
INFO - 9.mlp.down_proj 2ed error time: 0.002403736114501953
INFO - 9.mlp.down_proj proxy error after residual VPTQ: 1.688850517211904e-07, 1.7052871044143103e-05, 0.009903614409267902
INFO - ----Quantizing layer 10 ...---- 2026-02-23 09:36:07 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/10_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:36:08 10.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:10.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 66 iterations, error 8684.8984375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 10.self_attn.q_proj 1st kmeans time: 25.32708191871643
INFO - 10.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 10.self_attn.q_proj proxy error before VPTQ: 1.5782874243086553e-06, 0.00010997304343618453, 0.014351584948599339
INFO - 10.self_attn.q_proj 1st error time: 2.334010362625122
INFO - 10.self_attn.q_proj proxy error after VPTQ: 9.351193739348673e-07, 0.00010997304343618453, 0.008503168821334839
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3669.03271484375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 10.self_attn.q_proj residual time: 0.3150055408477783
INFO - 10.self_attn.q_proj 2ed gptq time: 2.6252524852752686
INFO - 10.self_attn.q_proj 2ed error time: 0.0003197193145751953
INFO - 10.self_attn.q_proj proxy error after residual VPTQ: 3.1747825346428726e-07, 0.00010997304343618453, 0.002886873437091708
INFO - load Hessian from /workspace/hessians/quip-format/hessians/10_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:36:40 10.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:10.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 25 iterations, error 1792.7994384765625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 10.self_attn.k_proj 1st kmeans time: 2.5437240600585938
INFO - 10.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 10.self_attn.k_proj proxy error before VPTQ: 4.2699051050476555e-07, 3.951299004256725e-05, 0.010806332342326641
INFO - 10.self_attn.k_proj 1st error time: 1.5449700355529785
INFO - 10.self_attn.k_proj proxy error after VPTQ: 3.087310744831484e-07, 3.951299004256725e-05, 0.007813407108187675
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 788.8240356445312
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 10.self_attn.k_proj residual time: 0.10043621063232422
INFO - 10.self_attn.k_proj 2ed gptq time: 2.0541152954101562
INFO - 10.self_attn.k_proj 2ed error time: 0.00037384033203125
INFO - 10.self_attn.k_proj proxy error after residual VPTQ: 1.0244925618962952e-07, 3.951299004256725e-05, 0.002592799486592412
INFO - load Hessian from /workspace/hessians/quip-format/hessians/10_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:36:48 10.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:10.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 34 iterations, error 2341.871826171875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 10.self_attn.v_proj 1st kmeans time: 3.387758493423462
INFO - 10.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 10.self_attn.v_proj proxy error before VPTQ: 1.2391463144467707e-07, 5.319883712218143e-06, 0.02329273335635662
INFO - 10.self_attn.v_proj 1st error time: 1.5484120845794678
INFO - 10.self_attn.v_proj proxy error after VPTQ: 8.636586557031478e-08, 5.319883712218143e-06, 0.016234539449214935
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 994.70947265625
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 10.self_attn.v_proj residual time: 0.09585881233215332
INFO - 10.self_attn.v_proj 2ed gptq time: 2.0642693042755127
INFO - 10.self_attn.v_proj 2ed error time: 0.0003762245178222656
INFO - 10.self_attn.v_proj proxy error after residual VPTQ: 2.7754690989922892e-08, 5.319883712218143e-06, 0.005217161029577255
INFO - load Hessian from /workspace/hessians/quip-format/hessians/10_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:36:56 10.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:10.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 78 iterations, error 18848.458984375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 10.self_attn.o_proj 1st kmeans time: 30.210407972335815
INFO - 10.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 10.self_attn.o_proj proxy error before VPTQ: 1.6168700085472665e-06, 3.8008805859135464e-05, 0.04253935441374779
INFO - 10.self_attn.o_proj 1st error time: 2.3106212615966797
INFO - 10.self_attn.o_proj proxy error after VPTQ: 9.075426987692481e-07, 3.8008805859135464e-05, 0.02387716993689537
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7303.71826171875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 10.self_attn.o_proj residual time: 0.3519458770751953
INFO - 10.self_attn.o_proj 2ed gptq time: 2.597074508666992
INFO - 10.self_attn.o_proj 2ed error time: 0.0010466575622558594
INFO - 10.self_attn.o_proj proxy error after residual VPTQ: 2.910145440182532e-07, 3.8008805859135464e-05, 0.00765650300309062
INFO - load Hessian from /workspace/hessians/quip-format/hessians/10_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:37:33 10.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:10.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 61756.7578125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 10.mlp.gate_proj 1st kmeans time: 132.40095162391663
INFO - 10.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 10.mlp.gate_proj proxy error before VPTQ: 9.50812864175532e-06, 0.00019899648032151163, 0.04778038710355759
INFO - 10.mlp.gate_proj 1st error time: 4.853024482727051
INFO - 10.mlp.gate_proj proxy error after VPTQ: 7.03659861756023e-06, 0.00019899648032151163, 0.035360418260097504
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 99 iterations, error 22874.46484375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 10.mlp.gate_proj residual time: 1.05607008934021
INFO - 10.mlp.gate_proj 2ed gptq time: 5.12499475479126
INFO - 10.mlp.gate_proj 2ed error time: 0.0012428760528564453
INFO - 10.mlp.gate_proj proxy error after residual VPTQ: 2.287699089720263e-06, 0.00019899648032151163, 0.011496178805828094
INFO - load Hessian from /workspace/hessians/quip-format/hessians/10_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:39:59 10.mlp.up_proj
INFO - enabling norm dim 0, layer_name:10.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 64622.4765625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 10.mlp.up_proj 1st kmeans time: 132.41028261184692
INFO - 10.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 10.mlp.up_proj proxy error before VPTQ: 7.870572517276742e-06, 0.00013955493341200054, 0.05639766529202461
INFO - 10.mlp.up_proj 1st error time: 4.7833075523376465
INFO - 10.mlp.up_proj proxy error after VPTQ: 5.70728707316448e-06, 0.00013955493341200054, 0.040896348655223846
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 99 iterations, error 23760.85546875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 10.mlp.up_proj residual time: 1.094994068145752
INFO - 10.mlp.up_proj 2ed gptq time: 5.1529271602630615
INFO - 10.mlp.up_proj 2ed error time: 0.0012881755828857422
INFO - 10.mlp.up_proj proxy error after residual VPTQ: 1.8436796835885616e-06, 0.00013955493341200054, 0.013211139477789402
INFO - load Hessian from /workspace/hessians/quip-format/hessians/10_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:42:27 10.mlp.down_proj
INFO - enabling norm dim 0, layer_name:10.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 85316.34375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 10.mlp.down_proj 1st kmeans time: 132.44386100769043
INFO - 10.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 10.mlp.down_proj proxy error before VPTQ: 6.970399795136473e-07, 1.7803356968215667e-05, 0.03915216401219368
INFO - 10.mlp.down_proj 1st error time: 8.548423528671265
INFO - 10.mlp.down_proj proxy error after VPTQ: 5.507561127160443e-07, 1.7803356968215667e-05, 0.03093552030622959
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 30087.267578125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 10.mlp.down_proj residual time: 1.1483922004699707
INFO - 10.mlp.down_proj 2ed gptq time: 9.503295421600342
INFO - 10.mlp.down_proj 2ed error time: 0.0032045841217041016
INFO - 10.mlp.down_proj proxy error after residual VPTQ: 1.7796182305573893e-07, 1.7803356968215667e-05, 0.009995969943702221
INFO - ----Quantizing layer 11 ...---- 2026-02-23 09:45:04 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/11_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:45:04 11.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:11.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 74 iterations, error 8235.37109375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 11.self_attn.q_proj 1st kmeans time: 28.486337184906006
INFO - 11.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 11.self_attn.q_proj proxy error before VPTQ: 1.654776951909298e-06, 8.478868403472006e-05, 0.019516482949256897
INFO - 11.self_attn.q_proj 1st error time: 2.263258457183838
INFO - 11.self_attn.q_proj proxy error after VPTQ: 1.0655852520358167e-06, 8.478868403472006e-05, 0.012567540630698204
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3479.45166015625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 11.self_attn.q_proj residual time: 0.3151857852935791
INFO - 11.self_attn.q_proj 2ed gptq time: 2.5564467906951904
INFO - 11.self_attn.q_proj 2ed error time: 0.0003769397735595703
INFO - 11.self_attn.q_proj proxy error after residual VPTQ: 3.6751674770130194e-07, 8.478868403472006e-05, 0.004334502387791872
INFO - load Hessian from /workspace/hessians/quip-format/hessians/11_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:45:40 11.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:11.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 25 iterations, error 1827.8046875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 11.self_attn.k_proj 1st kmeans time: 2.561671733856201
INFO - 11.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 11.self_attn.k_proj proxy error before VPTQ: 4.249856146998354e-07, 3.037096030311659e-05, 0.0139931570738554
INFO - 11.self_attn.k_proj 1st error time: 1.50447416305542
INFO - 11.self_attn.k_proj proxy error after VPTQ: 3.2446433806399e-07, 3.037096030311659e-05, 0.010683374479413033
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 808.613525390625
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 11.self_attn.k_proj residual time: 0.09039187431335449
INFO - 11.self_attn.k_proj 2ed gptq time: 1.9598469734191895
INFO - 11.self_attn.k_proj 2ed error time: 0.00037288665771484375
INFO - 11.self_attn.k_proj proxy error after residual VPTQ: 1.0930492067018349e-07, 3.037096030311659e-05, 0.0035989945754408836
INFO - load Hessian from /workspace/hessians/quip-format/hessians/11_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:45:47 11.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:11.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 21 iterations, error 2588.477783203125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 11.self_attn.v_proj 1st kmeans time: 2.169825792312622
INFO - 11.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 11.self_attn.v_proj proxy error before VPTQ: 1.574232726397895e-07, 7.23775747246691e-06, 0.021750282496213913
INFO - 11.self_attn.v_proj 1st error time: 1.493534803390503
INFO - 11.self_attn.v_proj proxy error after VPTQ: 1.1042008196682218e-07, 7.23775747246691e-06, 0.015256118029356003
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1085.0157470703125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 11.self_attn.v_proj residual time: 0.0890653133392334
INFO - 11.self_attn.v_proj 2ed gptq time: 1.9624836444854736
INFO - 11.self_attn.v_proj 2ed error time: 0.0003941059112548828
INFO - 11.self_attn.v_proj proxy error after residual VPTQ: 3.558992389685045e-08, 7.23775747246691e-06, 0.004917258396744728
INFO - load Hessian from /workspace/hessians/quip-format/hessians/11_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:45:54 11.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:11.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 69 iterations, error 17708.302734375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 11.self_attn.o_proj 1st kmeans time: 26.398451566696167
INFO - 11.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 11.self_attn.o_proj proxy error before VPTQ: 1.7436923371860757e-06, 4.330377123551443e-05, 0.040266525000333786
INFO - 11.self_attn.o_proj 1st error time: 2.2711870670318604
INFO - 11.self_attn.o_proj proxy error after VPTQ: 1.0869866855500732e-06, 4.330377123551443e-05, 0.02510143257677555
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 6865.28857421875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 11.self_attn.o_proj residual time: 0.3064417839050293
INFO - 11.self_attn.o_proj 2ed gptq time: 2.5618135929107666
INFO - 11.self_attn.o_proj 2ed error time: 0.0009987354278564453
INFO - 11.self_attn.o_proj proxy error after residual VPTQ: 3.5587839875006466e-07, 4.330377123551443e-05, 0.008218185044825077
INFO - load Hessian from /workspace/hessians/quip-format/hessians/11_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:46:28 11.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:11.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 59724.94140625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 11.mlp.gate_proj 1st kmeans time: 132.40598583221436
INFO - 11.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 11.mlp.gate_proj proxy error before VPTQ: 1.0130635928362608e-05, 0.00021235784515738487, 0.04770549386739731
INFO - 11.mlp.gate_proj 1st error time: 4.721880912780762
INFO - 11.mlp.gate_proj proxy error after VPTQ: 7.722050213487819e-06, 0.00021235784515738487, 0.036363385617733
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 22179.12890625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 11.mlp.gate_proj residual time: 1.0921132564544678
INFO - 11.mlp.gate_proj 2ed gptq time: 5.085377216339111
INFO - 11.mlp.gate_proj 2ed error time: 0.0012710094451904297
INFO - 11.mlp.gate_proj proxy error after residual VPTQ: 2.53247594628192e-06, 0.00021235784515738487, 0.011925511993467808
INFO - load Hessian from /workspace/hessians/quip-format/hessians/11_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:48:53 11.mlp.up_proj
INFO - enabling norm dim 0, layer_name:11.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 64076.453125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 11.mlp.up_proj 1st kmeans time: 132.40734887123108
INFO - 11.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 11.mlp.up_proj proxy error before VPTQ: 8.046028597163968e-06, 0.0001676385145401582, 0.04799630120396614
INFO - 11.mlp.up_proj 1st error time: 4.70813775062561
INFO - 11.mlp.up_proj proxy error after VPTQ: 5.842052814841736e-06, 0.0001676385145401582, 0.03484910726547241
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 90 iterations, error 23574.05078125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 11.mlp.up_proj residual time: 0.9814012050628662
INFO - 11.mlp.up_proj 2ed gptq time: 5.014710426330566
INFO - 11.mlp.up_proj 2ed error time: 0.001272439956665039
INFO - 11.mlp.up_proj proxy error after residual VPTQ: 1.8931016256829025e-06, 0.0001676385145401582, 0.011292760260403156
INFO - load Hessian from /workspace/hessians/quip-format/hessians/11_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:51:20 11.mlp.down_proj
INFO - enabling norm dim 0, layer_name:11.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 84576.375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 11.mlp.down_proj 1st kmeans time: 132.36464834213257
INFO - 11.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 11.mlp.down_proj proxy error before VPTQ: 7.058912956381391e-07, 1.6242212950601242e-05, 0.043460290879011154
INFO - 11.mlp.down_proj 1st error time: 8.43806767463684
INFO - 11.mlp.down_proj proxy error after VPTQ: 5.630968189507257e-07, 1.6242212950601242e-05, 0.03466872498393059
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 29786.62890625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 11.mlp.down_proj residual time: 1.0171535015106201
INFO - 11.mlp.down_proj 2ed gptq time: 9.480921983718872
INFO - 11.mlp.down_proj 2ed error time: 0.002295970916748047
INFO - 11.mlp.down_proj proxy error after residual VPTQ: 1.8228180920232262e-07, 1.6242212950601242e-05, 0.0112227201461792
INFO - ----Quantizing layer 12 ...---- 2026-02-23 09:53:56 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/12_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:53:57 12.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:12.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 69 iterations, error 7913.2255859375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 12.self_attn.q_proj 1st kmeans time: 26.605828523635864
INFO - 12.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 12.self_attn.q_proj proxy error before VPTQ: 1.6268056697299471e-06, 9.360098920296878e-05, 0.01738021895289421
INFO - 12.self_attn.q_proj 1st error time: 2.270885467529297
INFO - 12.self_attn.q_proj proxy error after VPTQ: 1.1153305194966379e-06, 9.360098920296878e-05, 0.011915798299014568
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3365.2578125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 12.self_attn.q_proj residual time: 0.3135192394256592
INFO - 12.self_attn.q_proj 2ed gptq time: 2.5936312675476074
INFO - 12.self_attn.q_proj 2ed error time: 0.0002841949462890625
INFO - 12.self_attn.q_proj proxy error after residual VPTQ: 3.890470736678253e-07, 9.360098920296878e-05, 0.004156441893428564
INFO - load Hessian from /workspace/hessians/quip-format/hessians/12_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:54:30 12.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:12.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 26 iterations, error 1710.1907958984375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 12.self_attn.k_proj 1st kmeans time: 2.638230562210083
INFO - 12.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 12.self_attn.k_proj proxy error before VPTQ: 4.0787858779367525e-07, 3.093975465162657e-05, 0.013182993978261948
INFO - 12.self_attn.k_proj 1st error time: 1.5159590244293213
INFO - 12.self_attn.k_proj proxy error after VPTQ: 3.1957685564520943e-07, 3.093975465162657e-05, 0.010329004377126694
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 759.70458984375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 12.self_attn.k_proj residual time: 0.08985424041748047
INFO - 12.self_attn.k_proj 2ed gptq time: 1.9573602676391602
INFO - 12.self_attn.k_proj 2ed error time: 0.0003364086151123047
INFO - 12.self_attn.k_proj proxy error after residual VPTQ: 1.0848437170807301e-07, 3.093975465162657e-05, 0.003506310051307082
INFO - load Hessian from /workspace/hessians/quip-format/hessians/12_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:54:38 12.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:12.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 23 iterations, error 2521.13232421875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 12.self_attn.v_proj 1st kmeans time: 2.371446132659912
INFO - 12.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 12.self_attn.v_proj proxy error before VPTQ: 1.3291892742017808e-07, 5.3912208386464044e-06, 0.024654699489474297
INFO - 12.self_attn.v_proj 1st error time: 1.5855700969696045
INFO - 12.self_attn.v_proj proxy error after VPTQ: 9.464518058166504e-08, 5.3912208386464044e-06, 0.01755542680621147
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1066.641357421875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 12.self_attn.v_proj residual time: 0.18917584419250488
INFO - 12.self_attn.v_proj 2ed gptq time: 1.956632375717163
INFO - 12.self_attn.v_proj 2ed error time: 0.0003578662872314453
INFO - 12.self_attn.v_proj proxy error after residual VPTQ: 3.0717018262294005e-08, 5.3912208386464044e-06, 0.005697599612176418
INFO - load Hessian from /workspace/hessians/quip-format/hessians/12_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:54:45 12.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:12.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 76 iterations, error 18300.302734375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 12.self_attn.o_proj 1st kmeans time: 29.02101230621338
INFO - 12.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 12.self_attn.o_proj proxy error before VPTQ: 1.4423239917960018e-06, 4.607867595041171e-05, 0.031301334500312805
INFO - 12.self_attn.o_proj 1st error time: 2.2684872150421143
INFO - 12.self_attn.o_proj proxy error after VPTQ: 8.530705599696375e-07, 4.607867595041171e-05, 0.018513347953557968
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7055.8271484375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 12.self_attn.o_proj residual time: 0.31108570098876953
INFO - 12.self_attn.o_proj 2ed gptq time: 2.5594637393951416
INFO - 12.self_attn.o_proj 2ed error time: 0.0009770393371582031
INFO - 12.self_attn.o_proj proxy error after residual VPTQ: 2.7505666366778314e-07, 4.607867595041171e-05, 0.0059692831709980965
INFO - load Hessian from /workspace/hessians/quip-format/hessians/12_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:55:21 12.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:12.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 62142.421875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 12.mlp.gate_proj 1st kmeans time: 132.41015338897705
INFO - 12.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 12.mlp.gate_proj proxy error before VPTQ: 9.684856195235625e-06, 0.00021561338508035988, 0.044917695224285126
INFO - 12.mlp.gate_proj 1st error time: 4.786856174468994
INFO - 12.mlp.gate_proj proxy error after VPTQ: 7.290150733751943e-06, 0.00021561338508035988, 0.033811215311288834
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 96 iterations, error 22853.5859375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 12.mlp.gate_proj residual time: 1.016439437866211
INFO - 12.mlp.gate_proj 2ed gptq time: 5.019555330276489
INFO - 12.mlp.gate_proj 2ed error time: 0.0011532306671142578
INFO - 12.mlp.gate_proj proxy error after residual VPTQ: 2.3716715986665804e-06, 0.00021561338508035988, 0.010999648831784725
INFO - load Hessian from /workspace/hessians/quip-format/hessians/12_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 09:57:47 12.mlp.up_proj
INFO - enabling norm dim 0, layer_name:12.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 65559.359375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 12.mlp.up_proj 1st kmeans time: 132.40370202064514
INFO - 12.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 12.mlp.up_proj proxy error before VPTQ: 8.196575436159037e-06, 0.0002209863596362993, 0.03709086775779724
INFO - 12.mlp.up_proj 1st error time: 4.7031683921813965
INFO - 12.mlp.up_proj proxy error after VPTQ: 6.0357765505614225e-06, 0.0002209863596362993, 0.02731289155781269
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 98 iterations, error 23950.826171875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 12.mlp.up_proj residual time: 0.9947311878204346
INFO - 12.mlp.up_proj 2ed gptq time: 5.014906883239746
INFO - 12.mlp.up_proj 2ed error time: 0.0010988712310791016
INFO - 12.mlp.up_proj proxy error after residual VPTQ: 1.953270384547068e-06, 0.0002209863596362993, 0.008838873356580734
INFO - load Hessian from /workspace/hessians/quip-format/hessians/12_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:00:14 12.mlp.down_proj
INFO - enabling norm dim 0, layer_name:12.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 84732.7265625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 12.mlp.down_proj 1st kmeans time: 132.379460811615
INFO - 12.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 12.mlp.down_proj proxy error before VPTQ: 6.805189514125232e-07, 1.4945539078325965e-05, 0.04553324729204178
INFO - 12.mlp.down_proj 1st error time: 8.426400423049927
INFO - 12.mlp.down_proj proxy error after VPTQ: 5.500733664121071e-07, 1.4945539078325965e-05, 0.03680518642067909
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 29808.943359375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 12.mlp.down_proj residual time: 1.0354869365692139
INFO - 12.mlp.down_proj 2ed gptq time: 9.52614140510559
INFO - 12.mlp.down_proj 2ed error time: 0.0023784637451171875
INFO - 12.mlp.down_proj proxy error after residual VPTQ: 1.782895395763262e-07, 1.4945539078325965e-05, 0.011929281055927277
INFO - ----Quantizing layer 13 ...---- 2026-02-23 10:02:50 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/13_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:02:51 13.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:13.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 70 iterations, error 10110.373046875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 13.self_attn.q_proj 1st kmeans time: 26.775662899017334
INFO - 13.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 13.self_attn.q_proj proxy error before VPTQ: 1.620263446966419e-06, 8.055223588598892e-05, 0.020114444196224213
INFO - 13.self_attn.q_proj 1st error time: 2.269012451171875
INFO - 13.self_attn.q_proj proxy error after VPTQ: 1.0919491160166217e-06, 8.055223588598892e-05, 0.013555789366364479
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 4145.39599609375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 13.self_attn.q_proj residual time: 0.32054591178894043
INFO - 13.self_attn.q_proj 2ed gptq time: 2.6177735328674316
INFO - 13.self_attn.q_proj 2ed error time: 0.00038623809814453125
INFO - 13.self_attn.q_proj proxy error after residual VPTQ: 3.67338060414113e-07, 8.055223588598892e-05, 0.0045602465979754925
INFO - load Hessian from /workspace/hessians/quip-format/hessians/13_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:03:25 13.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:13.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 21 iterations, error 2325.565185546875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 13.self_attn.k_proj 1st kmeans time: 2.174060821533203
INFO - 13.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 13.self_attn.k_proj proxy error before VPTQ: 4.5865664333177847e-07, 3.0376606446225196e-05, 0.015099008567631245
INFO - 13.self_attn.k_proj 1st error time: 1.4969615936279297
INFO - 13.self_attn.k_proj proxy error after VPTQ: 3.545308970842598e-07, 3.0376606446225196e-05, 0.01167118176817894
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 992.6453247070312
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 13.self_attn.k_proj residual time: 0.0885927677154541
INFO - 13.self_attn.k_proj 2ed gptq time: 1.9435651302337646
INFO - 13.self_attn.k_proj 2ed error time: 0.00041484832763671875
INFO - 13.self_attn.k_proj proxy error after residual VPTQ: 1.1640990749128832e-07, 3.0376606446225196e-05, 0.003832222195342183
INFO - load Hessian from /workspace/hessians/quip-format/hessians/13_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:03:32 13.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:13.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 23 iterations, error 2892.15087890625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 13.self_attn.v_proj 1st kmeans time: 2.3535659313201904
INFO - 13.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 13.self_attn.v_proj proxy error before VPTQ: 1.5948457132708427e-07, 5.954037987976335e-06, 0.026785951107740402
INFO - 13.self_attn.v_proj 1st error time: 1.4900095462799072
INFO - 13.self_attn.v_proj proxy error after VPTQ: 1.1829777690763876e-07, 5.954037987976335e-06, 0.01986849494278431
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1203.229248046875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 13.self_attn.v_proj residual time: 0.08978462219238281
INFO - 13.self_attn.v_proj 2ed gptq time: 1.9415802955627441
INFO - 13.self_attn.v_proj 2ed error time: 0.00040435791015625
INFO - 13.self_attn.v_proj proxy error after residual VPTQ: 3.7622736925868594e-08, 5.954037987976335e-06, 0.006318860687315464
INFO - load Hessian from /workspace/hessians/quip-format/hessians/13_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:03:39 13.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:13.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 81 iterations, error 19723.708984375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 13.self_attn.o_proj 1st kmeans time: 31.23781156539917
INFO - 13.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 13.self_attn.o_proj proxy error before VPTQ: 2.0215616132190917e-06, 4.5905278966529295e-05, 0.044037673622369766
INFO - 13.self_attn.o_proj 1st error time: 2.2811341285705566
INFO - 13.self_attn.o_proj proxy error after VPTQ: 1.2224522834003437e-06, 4.5905278966529295e-05, 0.02662988379597664
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7471.84326171875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 13.self_attn.o_proj residual time: 0.332486629486084
INFO - 13.self_attn.o_proj 2ed gptq time: 2.5495264530181885
INFO - 13.self_attn.o_proj 2ed error time: 0.0011053085327148438
INFO - 13.self_attn.o_proj proxy error after residual VPTQ: 3.93987193092471e-07, 4.5905278966529295e-05, 0.008582611568272114
INFO - load Hessian from /workspace/hessians/quip-format/hessians/13_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:04:17 13.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:13.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 61471.56640625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 13.mlp.gate_proj 1st kmeans time: 132.40992951393127
INFO - 13.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 13.mlp.gate_proj proxy error before VPTQ: 9.664314347901382e-06, 0.00018441352585796267, 0.052405670285224915
INFO - 13.mlp.gate_proj 1st error time: 4.722787857055664
INFO - 13.mlp.gate_proj proxy error after VPTQ: 7.49917217035545e-06, 0.00018441352585796267, 0.040664978325366974
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 96 iterations, error 22608.484375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 13.mlp.gate_proj residual time: 0.980903148651123
INFO - 13.mlp.gate_proj 2ed gptq time: 5.0941479206085205
INFO - 13.mlp.gate_proj 2ed error time: 0.0014760494232177734
INFO - 13.mlp.gate_proj proxy error after residual VPTQ: 2.448067334626103e-06, 0.00018441352585796267, 0.013274879194796085
INFO - load Hessian from /workspace/hessians/quip-format/hessians/13_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:06:43 13.mlp.up_proj
INFO - enabling norm dim 0, layer_name:13.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 65364.5625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 13.mlp.up_proj 1st kmeans time: 132.41877794265747
INFO - 13.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 13.mlp.up_proj proxy error before VPTQ: 8.164856808434706e-06, 0.00014387564442586154, 0.05674940347671509
INFO - 13.mlp.up_proj 1st error time: 4.74458122253418
INFO - 13.mlp.up_proj proxy error after VPTQ: 6.147590283944737e-06, 0.00014387564442586154, 0.042728498578071594
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 23844.220703125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 13.mlp.up_proj residual time: 1.149456262588501
INFO - 13.mlp.up_proj 2ed gptq time: 5.057386875152588
INFO - 13.mlp.up_proj 2ed error time: 0.0012142658233642578
INFO - 13.mlp.up_proj proxy error after residual VPTQ: 1.989220891118748e-06, 0.00014387564442586154, 0.013825973495841026
INFO - load Hessian from /workspace/hessians/quip-format/hessians/13_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:09:10 13.mlp.down_proj
INFO - enabling norm dim 0, layer_name:13.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 84246.9765625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 13.mlp.down_proj 1st kmeans time: 132.4647138118744
INFO - 13.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 13.mlp.down_proj proxy error before VPTQ: 6.908725254106685e-07, 1.4973616089264397e-05, 0.046139322221279144
INFO - 13.mlp.down_proj 1st error time: 8.65036654472351
INFO - 13.mlp.down_proj proxy error after VPTQ: 5.739509560953593e-07, 1.4973616089264397e-05, 0.038330819457769394
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 29522.720703125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 13.mlp.down_proj residual time: 1.0142440795898438
INFO - 13.mlp.down_proj 2ed gptq time: 9.629350185394287
INFO - 13.mlp.down_proj 2ed error time: 0.002599477767944336
INFO - 13.mlp.down_proj proxy error after residual VPTQ: 1.8664390211142745e-07, 1.4973616089264397e-05, 0.012464851140975952
INFO - ----Quantizing layer 14 ...---- 2026-02-23 10:11:47 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/14_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:11:47 14.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:14.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 73 iterations, error 8741.8447265625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 14.self_attn.q_proj 1st kmeans time: 28.18520712852478
INFO - 14.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 14.self_attn.q_proj proxy error before VPTQ: 1.5553540606561e-06, 7.212701893877238e-05, 0.021564098075032234
INFO - 14.self_attn.q_proj 1st error time: 2.287550687789917
INFO - 14.self_attn.q_proj proxy error after VPTQ: 1.0893621720242663e-06, 7.212701893877238e-05, 0.015103385783731937
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3672.2919921875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 14.self_attn.q_proj residual time: 0.3175380229949951
INFO - 14.self_attn.q_proj 2ed gptq time: 2.603311538696289
INFO - 14.self_attn.q_proj 2ed error time: 0.0004284381866455078
INFO - 14.self_attn.q_proj proxy error after residual VPTQ: 3.785516469179129e-07, 7.212701893877238e-05, 0.005248402711004019
INFO - load Hessian from /workspace/hessians/quip-format/hessians/14_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:12:23 14.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:14.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 28 iterations, error 2084.51123046875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 14.self_attn.k_proj 1st kmeans time: 2.82902455329895
INFO - 14.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 14.self_attn.k_proj proxy error before VPTQ: 3.53253398088782e-07, 1.8663158698473126e-05, 0.01892784610390663
INFO - 14.self_attn.k_proj 1st error time: 1.5345160961151123
INFO - 14.self_attn.k_proj proxy error after VPTQ: 2.8146342856416595e-07, 1.8663158698473126e-05, 0.015081232413649559
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 903.7887573242188
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 14.self_attn.k_proj residual time: 0.09360980987548828
INFO - 14.self_attn.k_proj 2ed gptq time: 2.044503927230835
INFO - 14.self_attn.k_proj 2ed error time: 0.0004973411560058594
INFO - 14.self_attn.k_proj proxy error after residual VPTQ: 9.516992349745124e-08, 1.8663158698473126e-05, 0.005099346861243248
INFO - load Hessian from /workspace/hessians/quip-format/hessians/14_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:12:31 14.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:14.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 2396.70703125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 14.self_attn.v_proj 1st kmeans time: 2.27781343460083
INFO - 14.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 14.self_attn.v_proj proxy error before VPTQ: 2.074454243938817e-07, 7.72451312514022e-06, 0.02685546875
INFO - 14.self_attn.v_proj 1st error time: 1.5931780338287354
INFO - 14.self_attn.v_proj proxy error after VPTQ: 1.6269076752450928e-07, 7.72451312514022e-06, 0.021061621606349945
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1036.1561279296875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 14.self_attn.v_proj residual time: 0.09665274620056152
INFO - 14.self_attn.v_proj 2ed gptq time: 2.074843406677246
INFO - 14.self_attn.v_proj 2ed error time: 0.0006368160247802734
INFO - 14.self_attn.v_proj proxy error after residual VPTQ: 5.3920182097044744e-08, 7.72451312514022e-06, 0.006980398669838905
INFO - load Hessian from /workspace/hessians/quip-format/hessians/14_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:12:38 14.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:14.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 61 iterations, error 18034.1328125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 14.self_attn.o_proj 1st kmeans time: 23.484543800354004
INFO - 14.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 14.self_attn.o_proj proxy error before VPTQ: 1.955676907527959e-06, 8.133801748044789e-05, 0.024043822661042213
INFO - 14.self_attn.o_proj 1st error time: 2.4007861614227295
INFO - 14.self_attn.o_proj proxy error after VPTQ: 1.1935312613786664e-06, 8.133801748044789e-05, 0.014673719182610512
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 6875.79296875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 14.self_attn.o_proj residual time: 0.3153562545776367
INFO - 14.self_attn.o_proj 2ed gptq time: 2.5964903831481934
INFO - 14.self_attn.o_proj 2ed error time: 0.001165628433227539
INFO - 14.self_attn.o_proj proxy error after residual VPTQ: 3.9144120478340483e-07, 8.133801748044789e-05, 0.004812524653971195
INFO - load Hessian from /workspace/hessians/quip-format/hessians/14_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:13:09 14.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:14.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 56183.73828125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 14.mlp.gate_proj 1st kmeans time: 132.50375699996948
INFO - 14.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 14.mlp.gate_proj proxy error before VPTQ: 1.0243677024845965e-05, 0.00023659302678424865, 0.04329661279916763
INFO - 14.mlp.gate_proj 1st error time: 4.785433053970337
INFO - 14.mlp.gate_proj proxy error after VPTQ: 8.14858140074648e-06, 0.00023659302678424865, 0.03444134071469307
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 91 iterations, error 20787.44140625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 14.mlp.gate_proj residual time: 1.0203239917755127
INFO - 14.mlp.gate_proj 2ed gptq time: 5.06178879737854
INFO - 14.mlp.gate_proj 2ed error time: 0.0015010833740234375
INFO - 14.mlp.gate_proj proxy error after residual VPTQ: 2.7004973617295036e-06, 0.00023659302678424865, 0.011414104141294956
INFO - load Hessian from /workspace/hessians/quip-format/hessians/14_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:15:35 14.mlp.up_proj
INFO - enabling norm dim 0, layer_name:14.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 63114.38671875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 14.mlp.up_proj 1st kmeans time: 132.42747449874878
INFO - 14.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 14.mlp.up_proj proxy error before VPTQ: 8.040124157560058e-06, 0.00021703843958675861, 0.0370447002351284
INFO - 14.mlp.up_proj 1st error time: 4.811326265335083
INFO - 14.mlp.up_proj proxy error after VPTQ: 6.097204732213868e-06, 0.00021703843958675861, 0.02809274196624756
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 93 iterations, error 23088.77734375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 14.mlp.up_proj residual time: 0.9787464141845703
INFO - 14.mlp.up_proj 2ed gptq time: 5.072679042816162
INFO - 14.mlp.up_proj 2ed error time: 0.0014066696166992188
INFO - 14.mlp.up_proj proxy error after residual VPTQ: 1.9949957277276553e-06, 0.00021703843958675861, 0.009191900491714478
INFO - load Hessian from /workspace/hessians/quip-format/hessians/14_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:18:03 14.mlp.down_proj
INFO - enabling norm dim 0, layer_name:14.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 82733.171875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 14.mlp.down_proj 1st kmeans time: 132.3999412059784
INFO - 14.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 14.mlp.down_proj proxy error before VPTQ: 7.160895734159567e-07, 2.068704088742379e-05, 0.034615371376276016
INFO - 14.mlp.down_proj 1st error time: 8.639461517333984
INFO - 14.mlp.down_proj proxy error after VPTQ: 5.816830253024818e-07, 2.068704088742379e-05, 0.028118232265114784
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 29209.62109375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 14.mlp.down_proj residual time: 1.016796350479126
INFO - 14.mlp.down_proj 2ed gptq time: 9.632542371749878
INFO - 14.mlp.down_proj 2ed error time: 0.002517223358154297
INFO - 14.mlp.down_proj proxy error after residual VPTQ: 1.896655135169567e-07, 2.068704088742379e-05, 0.009168324992060661
INFO - ----Quantizing layer 15 ...---- 2026-02-23 10:20:39 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/15_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:20:39 15.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:15.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 90 iterations, error 9430.361328125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 15.self_attn.q_proj 1st kmeans time: 34.291186571121216
INFO - 15.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 15.self_attn.q_proj proxy error before VPTQ: 1.6847545794007601e-06, 7.219465624075383e-05, 0.023336278274655342
INFO - 15.self_attn.q_proj 1st error time: 2.3298535346984863
INFO - 15.self_attn.q_proj proxy error after VPTQ: 1.2097632406948833e-06, 7.219465624075383e-05, 0.01675696298480034
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3953.0908203125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 15.self_attn.q_proj residual time: 0.3134605884552002
INFO - 15.self_attn.q_proj 2ed gptq time: 2.601811408996582
INFO - 15.self_attn.q_proj 2ed error time: 0.00030684471130371094
INFO - 15.self_attn.q_proj proxy error after residual VPTQ: 4.1779412640607916e-07, 7.219465624075383e-05, 0.005787050351500511
INFO - load Hessian from /workspace/hessians/quip-format/hessians/15_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:21:21 15.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:15.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 25 iterations, error 2145.032958984375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 15.self_attn.k_proj 1st kmeans time: 2.551919937133789
INFO - 15.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 15.self_attn.k_proj proxy error before VPTQ: 3.8714910033377237e-07, 1.970171979337465e-05, 0.019650522619485855
INFO - 15.self_attn.k_proj 1st error time: 1.5542104244232178
INFO - 15.self_attn.k_proj proxy error after VPTQ: 3.197003195509751e-07, 1.970171979337465e-05, 0.016227025538682938
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 938.892333984375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 15.self_attn.k_proj residual time: 0.09664630889892578
INFO - 15.self_attn.k_proj 2ed gptq time: 2.0635201930999756
INFO - 15.self_attn.k_proj 2ed error time: 0.0005369186401367188
INFO - 15.self_attn.k_proj proxy error after residual VPTQ: 1.0828293284248502e-07, 1.970171979337465e-05, 0.005496115889400244
INFO - load Hessian from /workspace/hessians/quip-format/hessians/15_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:21:28 15.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:15.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 26 iterations, error 2941.540283203125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 15.self_attn.v_proj 1st kmeans time: 2.640714168548584
INFO - 15.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 15.self_attn.v_proj proxy error before VPTQ: 1.8869604900828563e-07, 5.879370291950181e-06, 0.032094601541757584
INFO - 15.self_attn.v_proj 1st error time: 1.536832571029663
INFO - 15.self_attn.v_proj proxy error after VPTQ: 1.3991422065373627e-07, 5.879370291950181e-06, 0.02379748411476612
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1196.782470703125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 15.self_attn.v_proj residual time: 0.09035801887512207
INFO - 15.self_attn.v_proj 2ed gptq time: 2.0381052494049072
INFO - 15.self_attn.v_proj 2ed error time: 0.0003936290740966797
INFO - 15.self_attn.v_proj proxy error after residual VPTQ: 4.551827714749379e-08, 5.879370291950181e-06, 0.007742032874375582
INFO - load Hessian from /workspace/hessians/quip-format/hessians/15_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:21:36 15.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:15.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 67 iterations, error 20378.91015625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 15.self_attn.o_proj 1st kmeans time: 25.665815114974976
INFO - 15.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 15.self_attn.o_proj proxy error before VPTQ: 1.8221657001049607e-06, 9.244457760360092e-05, 0.019710898399353027
INFO - 15.self_attn.o_proj 1st error time: 2.3326358795166016
INFO - 15.self_attn.o_proj proxy error after VPTQ: 1.0404321528767468e-06, 9.244457760360092e-05, 0.011254658922553062
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7663.59814453125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 15.self_attn.o_proj residual time: 0.3152952194213867
INFO - 15.self_attn.o_proj 2ed gptq time: 2.593775510787964
INFO - 15.self_attn.o_proj 2ed error time: 0.001056671142578125
INFO - 15.self_attn.o_proj proxy error after residual VPTQ: 3.307222868897952e-07, 9.244457760360092e-05, 0.0035775196738541126
INFO - load Hessian from /workspace/hessians/quip-format/hessians/15_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:22:09 15.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:15.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 58030.515625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 15.mlp.gate_proj 1st kmeans time: 132.41091871261597
INFO - 15.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 15.mlp.gate_proj proxy error before VPTQ: 1.122019239119254e-05, 0.0003304706478957087, 0.03395216166973114
INFO - 15.mlp.gate_proj 1st error time: 4.763918161392212
INFO - 15.mlp.gate_proj proxy error after VPTQ: 9.042340934684034e-06, 0.0003304706478957087, 0.02736200951039791
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 100 iterations, error 21398.5703125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 15.mlp.gate_proj residual time: 1.0448215007781982
INFO - 15.mlp.gate_proj 2ed gptq time: 5.0545594692230225
INFO - 15.mlp.gate_proj 2ed error time: 0.0013513565063476562
INFO - 15.mlp.gate_proj proxy error after residual VPTQ: 2.998440777446376e-06, 0.0003304706478957087, 0.009073243476450443
INFO - load Hessian from /workspace/hessians/quip-format/hessians/15_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:24:35 15.mlp.up_proj
INFO - enabling norm dim 0, layer_name:15.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 66209.609375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 15.mlp.up_proj 1st kmeans time: 132.40584635734558
INFO - 15.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 15.mlp.up_proj proxy error before VPTQ: 7.655893568880856e-06, 0.00012635477469302714, 0.060590457171201706
INFO - 15.mlp.up_proj 1st error time: 4.742987155914307
INFO - 15.mlp.up_proj proxy error after VPTQ: 5.725327355321497e-06, 0.00012635477469302714, 0.045311521738767624
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 90 iterations, error 24027.00390625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 15.mlp.up_proj residual time: 0.9399290084838867
INFO - 15.mlp.up_proj 2ed gptq time: 5.132649660110474
INFO - 15.mlp.up_proj 2ed error time: 0.0011873245239257812
INFO - 15.mlp.up_proj proxy error after residual VPTQ: 1.8595335404825164e-06, 0.00012635477469302714, 0.01471676491200924
INFO - load Hessian from /workspace/hessians/quip-format/hessians/15_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:27:02 15.mlp.down_proj
INFO - enabling norm dim 0, layer_name:15.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 85822.1171875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 15.mlp.down_proj 1st kmeans time: 132.36821246147156
INFO - 15.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 15.mlp.down_proj proxy error before VPTQ: 6.720952114847023e-07, 1.5727779100416228e-05, 0.042733002454042435
INFO - 15.mlp.down_proj 1st error time: 8.527230024337769
INFO - 15.mlp.down_proj proxy error after VPTQ: 5.368827942220378e-07, 1.5727779100416228e-05, 0.034135956317186356
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 30056.509765625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 15.mlp.down_proj residual time: 1.02024245262146
INFO - 15.mlp.down_proj 2ed gptq time: 9.60785961151123
INFO - 15.mlp.down_proj 2ed error time: 0.0024442672729492188
INFO - 15.mlp.down_proj proxy error after residual VPTQ: 1.746271749425432e-07, 1.5727779100416228e-05, 0.011103104799985886
INFO - ----Quantizing layer 16 ...---- 2026-02-23 10:29:38 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/16_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:29:38 16.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:16.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 77 iterations, error 10555.861328125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 16.self_attn.q_proj 1st kmeans time: 29.373651266098022
INFO - 16.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 16.self_attn.q_proj proxy error before VPTQ: 1.6628357570880326e-06, 7.007057865848765e-05, 0.023730870336294174
INFO - 16.self_attn.q_proj 1st error time: 2.364316940307617
INFO - 16.self_attn.q_proj proxy error after VPTQ: 1.1517424809426302e-06, 7.007057865848765e-05, 0.01643689163029194
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 4352.28515625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 16.self_attn.q_proj residual time: 0.3147759437561035
INFO - 16.self_attn.q_proj 2ed gptq time: 2.5928683280944824
INFO - 16.self_attn.q_proj 2ed error time: 0.00034046173095703125
INFO - 16.self_attn.q_proj proxy error after residual VPTQ: 3.9117958294809796e-07, 7.007057865848765e-05, 0.005582651123404503
INFO - load Hessian from /workspace/hessians/quip-format/hessians/16_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:30:15 16.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:16.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 24 iterations, error 2478.865966796875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 16.self_attn.k_proj 1st kmeans time: 2.46779727935791
INFO - 16.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 16.self_attn.k_proj proxy error before VPTQ: 4.215376918637048e-07, 2.1019179257564247e-05, 0.020054906606674194
INFO - 16.self_attn.k_proj 1st error time: 1.5264360904693604
INFO - 16.self_attn.k_proj proxy error after VPTQ: 3.367956651345594e-07, 2.1019179257564247e-05, 0.01602325402200222
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1059.9990234375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 16.self_attn.k_proj residual time: 0.09656119346618652
INFO - 16.self_attn.k_proj 2ed gptq time: 2.1212868690490723
INFO - 16.self_attn.k_proj 2ed error time: 0.0004153251647949219
INFO - 16.self_attn.k_proj proxy error after residual VPTQ: 1.1131056254498617e-07, 2.1019179257564247e-05, 0.00529566640034318
INFO - load Hessian from /workspace/hessians/quip-format/hessians/16_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:30:22 16.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:16.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 3064.134765625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 16.self_attn.v_proj 1st kmeans time: 2.320358991622925
INFO - 16.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 16.self_attn.v_proj proxy error before VPTQ: 1.8963123693538364e-07, 5.447644525702344e-06, 0.034809768199920654
INFO - 16.self_attn.v_proj 1st error time: 1.529677152633667
INFO - 16.self_attn.v_proj proxy error after VPTQ: 1.4397414815903176e-07, 5.447644525702344e-06, 0.026428697630763054
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1271.6043701171875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 16.self_attn.v_proj residual time: 0.09454798698425293
INFO - 16.self_attn.v_proj 2ed gptq time: 2.042908191680908
INFO - 16.self_attn.v_proj 2ed error time: 0.0004298686981201172
INFO - 16.self_attn.v_proj proxy error after residual VPTQ: 4.6402995224070764e-08, 5.447644525702344e-06, 0.008517992682754993
INFO - load Hessian from /workspace/hessians/quip-format/hessians/16_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:30:30 16.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:16.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 62 iterations, error 20261.57421875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 16.self_attn.o_proj 1st kmeans time: 23.991304874420166
INFO - 16.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 16.self_attn.o_proj proxy error before VPTQ: 2.2978517790761543e-06, 0.00020604624296538532, 0.0111521165817976
INFO - 16.self_attn.o_proj 1st error time: 2.346008539199829
INFO - 16.self_attn.o_proj proxy error after VPTQ: 1.3734861568082124e-06, 0.00020604624296538532, 0.006665911991149187
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7594.09033203125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 16.self_attn.o_proj residual time: 0.3171272277832031
INFO - 16.self_attn.o_proj 2ed gptq time: 2.5987606048583984
INFO - 16.self_attn.o_proj 2ed error time: 0.001088857650756836
INFO - 16.self_attn.o_proj proxy error after residual VPTQ: 4.442930503500975e-07, 0.00020604624296538532, 0.0021562783513218164
INFO - load Hessian from /workspace/hessians/quip-format/hessians/16_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:31:01 16.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:16.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 53682.203125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 16.mlp.gate_proj 1st kmeans time: 133.0082459449768
INFO - 16.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 16.mlp.gate_proj proxy error before VPTQ: 1.1250129318796098e-05, 0.000296480895485729, 0.037945546209812164
INFO - 16.mlp.gate_proj 1st error time: 4.771161794662476
INFO - 16.mlp.gate_proj proxy error after VPTQ: 9.05276738194516e-06, 0.000296480895485729, 0.030534066259860992
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 96 iterations, error 19930.0
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 16.mlp.gate_proj residual time: 1.0343515872955322
INFO - 16.mlp.gate_proj 2ed gptq time: 5.052999973297119
INFO - 16.mlp.gate_proj 2ed error time: 0.0013074874877929688
INFO - 16.mlp.gate_proj proxy error after residual VPTQ: 3.0234054975153413e-06, 0.000296480895485729, 0.010197640396654606
INFO - load Hessian from /workspace/hessians/quip-format/hessians/16_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:33:27 16.mlp.up_proj
INFO - enabling norm dim 0, layer_name:16.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 65335.69140625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 16.mlp.up_proj 1st kmeans time: 132.61236310005188
INFO - 16.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 16.mlp.up_proj proxy error before VPTQ: 7.458035724994261e-06, 0.0001726532937027514, 0.043196603655815125
INFO - 16.mlp.up_proj 1st error time: 4.741183042526245
INFO - 16.mlp.up_proj proxy error after VPTQ: 5.5596442507521715e-06, 0.0001726532937027514, 0.03220120444893837
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 96 iterations, error 23734.640625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 16.mlp.up_proj residual time: 1.051497459411621
INFO - 16.mlp.up_proj 2ed gptq time: 5.049030780792236
INFO - 16.mlp.up_proj 2ed error time: 0.0011124610900878906
INFO - 16.mlp.up_proj proxy error after residual VPTQ: 1.8065356925944798e-06, 0.0001726532937027514, 0.010463371872901917
INFO - load Hessian from /workspace/hessians/quip-format/hessians/16_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:35:55 16.mlp.down_proj
INFO - enabling norm dim 0, layer_name:16.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 85400.9296875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 16.mlp.down_proj 1st kmeans time: 132.3801212310791
INFO - 16.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 16.mlp.down_proj proxy error before VPTQ: 6.47939998543734e-07, 1.777715442585759e-05, 0.03644790127873421
INFO - 16.mlp.down_proj 1st error time: 8.801833629608154
INFO - 16.mlp.down_proj proxy error after VPTQ: 5.091294497105991e-07, 1.777715442585759e-05, 0.0286395363509655
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 30001.32421875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 16.mlp.down_proj residual time: 1.0603365898132324
INFO - 16.mlp.down_proj 2ed gptq time: 9.629889011383057
INFO - 16.mlp.down_proj 2ed error time: 0.0024576187133789062
INFO - 16.mlp.down_proj proxy error after residual VPTQ: 1.65183237754718e-07, 1.777715442585759e-05, 0.009291882626712322
INFO - ----Quantizing layer 17 ...---- 2026-02-23 10:38:31 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/17_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:38:31 17.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:17.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 81 iterations, error 9373.783203125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 17.self_attn.q_proj 1st kmeans time: 30.8843252658844
INFO - 17.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 17.self_attn.q_proj proxy error before VPTQ: 1.5718185295554576e-06, 9.325920109404251e-05, 0.0168542992323637
INFO - 17.self_attn.q_proj 1st error time: 2.3516080379486084
INFO - 17.self_attn.q_proj proxy error after VPTQ: 1.047745399773703e-06, 9.325920109404251e-05, 0.01123476680368185
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3867.843505859375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 17.self_attn.q_proj residual time: 0.31999778747558594
INFO - 17.self_attn.q_proj 2ed gptq time: 2.588526964187622
INFO - 17.self_attn.q_proj 2ed error time: 0.0002808570861816406
INFO - 17.self_attn.q_proj proxy error after residual VPTQ: 3.5893248195861815e-07, 9.325920109404251e-05, 0.0038487622514367104
INFO - load Hessian from /workspace/hessians/quip-format/hessians/17_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:39:09 17.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:17.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 28 iterations, error 2073.21044921875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 17.self_attn.k_proj 1st kmeans time: 2.8333280086517334
INFO - 17.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 17.self_attn.k_proj proxy error before VPTQ: 3.7122279650247947e-07, 2.2861186153022572e-05, 0.016238125041127205
INFO - 17.self_attn.k_proj 1st error time: 1.54939866065979
INFO - 17.self_attn.k_proj proxy error after VPTQ: 2.9012727509325487e-07, 2.2861186153022572e-05, 0.012690823525190353
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 893.0223388671875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 17.self_attn.k_proj residual time: 0.16240739822387695
INFO - 17.self_attn.k_proj 2ed gptq time: 2.0622551441192627
INFO - 17.self_attn.k_proj 2ed error time: 0.00036787986755371094
INFO - 17.self_attn.k_proj proxy error after residual VPTQ: 9.684546853350184e-08, 2.2861186153022572e-05, 0.004236239939928055
INFO - load Hessian from /workspace/hessians/quip-format/hessians/17_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:39:17 17.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:17.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 29 iterations, error 2499.737548828125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 17.self_attn.v_proj 1st kmeans time: 2.930797576904297
INFO - 17.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 17.self_attn.v_proj proxy error before VPTQ: 1.637450566249754e-07, 5.965476702840533e-06, 0.027448778972029686
INFO - 17.self_attn.v_proj 1st error time: 1.5305864810943604
INFO - 17.self_attn.v_proj proxy error after VPTQ: 1.2543121385988343e-07, 5.965476702840533e-06, 0.021026184782385826
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1056.51611328125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 17.self_attn.v_proj residual time: 0.09076642990112305
INFO - 17.self_attn.v_proj 2ed gptq time: 2.063908338546753
INFO - 17.self_attn.v_proj 2ed error time: 0.0003981590270996094
INFO - 17.self_attn.v_proj proxy error after residual VPTQ: 4.102843931264033e-08, 5.965476702840533e-06, 0.006877646315842867
INFO - load Hessian from /workspace/hessians/quip-format/hessians/17_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:39:25 17.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:17.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 71 iterations, error 19490.5390625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 17.self_attn.o_proj 1st kmeans time: 27.168219804763794
INFO - 17.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 17.self_attn.o_proj proxy error before VPTQ: 2.1522605493373703e-06, 0.00011554095544852316, 0.01862768456339836
INFO - 17.self_attn.o_proj 1st error time: 2.311535596847534
INFO - 17.self_attn.o_proj proxy error after VPTQ: 1.4319102774607018e-06, 0.00011554095544852316, 0.012393097393214703
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 86 iterations, error 7450.46728515625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 17.self_attn.o_proj residual time: 0.27763843536376953
INFO - 17.self_attn.o_proj 2ed gptq time: 2.5950303077697754
INFO - 17.self_attn.o_proj 2ed error time: 0.0009899139404296875
INFO - 17.self_attn.o_proj proxy error after residual VPTQ: 4.6203584247450635e-07, 0.00011554095544852316, 0.0039988923817873
INFO - load Hessian from /workspace/hessians/quip-format/hessians/17_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:39:59 17.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:17.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 51570.390625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 17.mlp.gate_proj 1st kmeans time: 132.40162634849548
INFO - 17.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 17.mlp.gate_proj proxy error before VPTQ: 1.106747004087083e-05, 0.00027457132819108665, 0.04030817747116089
INFO - 17.mlp.gate_proj 1st error time: 4.758021593093872
INFO - 17.mlp.gate_proj proxy error after VPTQ: 8.849807272781618e-06, 0.00027457132819108665, 0.03223136067390442
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 97 iterations, error 19092.142578125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 17.mlp.gate_proj residual time: 1.0472936630249023
INFO - 17.mlp.gate_proj 2ed gptq time: 5.0550267696380615
INFO - 17.mlp.gate_proj 2ed error time: 0.001215219497680664
INFO - 17.mlp.gate_proj proxy error after residual VPTQ: 2.9583570722024888e-06, 0.00027457132819108665, 0.010774457827210426
INFO - load Hessian from /workspace/hessians/quip-format/hessians/17_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:42:25 17.mlp.up_proj
INFO - enabling norm dim 0, layer_name:17.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 66599.9375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 17.mlp.up_proj 1st kmeans time: 132.4061758518219
INFO - 17.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 17.mlp.up_proj proxy error before VPTQ: 7.626887963851914e-06, 0.0001280773722101003, 0.05954906716942787
INFO - 17.mlp.up_proj 1st error time: 4.820184707641602
INFO - 17.mlp.up_proj proxy error after VPTQ: 5.836728632857557e-06, 0.0001280773722101003, 0.04557189717888832
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 24038.333984375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 17.mlp.up_proj residual time: 1.042827844619751
INFO - 17.mlp.up_proj 2ed gptq time: 5.072084665298462
INFO - 17.mlp.up_proj 2ed error time: 0.0011851787567138672
INFO - 17.mlp.up_proj proxy error after residual VPTQ: 1.8915225155069493e-06, 0.0001280773722101003, 0.014768592081964016
INFO - load Hessian from /workspace/hessians/quip-format/hessians/17_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:44:52 17.mlp.down_proj
INFO - enabling norm dim 0, layer_name:17.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 85239.0
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 17.mlp.down_proj 1st kmeans time: 132.39766669273376
INFO - 17.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 17.mlp.down_proj proxy error before VPTQ: 6.576871101060533e-07, 1.6744112144806422e-05, 0.03927870839834213
INFO - 17.mlp.down_proj 1st error time: 8.866026163101196
INFO - 17.mlp.down_proj proxy error after VPTQ: 5.305519152898341e-07, 1.6744112144806422e-05, 0.031685877591371536
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 29792.251953125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 17.mlp.down_proj residual time: 1.0223603248596191
INFO - 17.mlp.down_proj 2ed gptq time: 9.867780923843384
INFO - 17.mlp.down_proj 2ed error time: 0.0027959346771240234
INFO - 17.mlp.down_proj proxy error after residual VPTQ: 1.7210297187375545e-07, 1.6744112144806422e-05, 0.010278417728841305
INFO - ----Quantizing layer 18 ...---- 2026-02-23 10:47:29 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/18_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:47:30 18.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:18.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 79 iterations, error 8237.8037109375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 18.self_attn.q_proj 1st kmeans time: 30.116331100463867
INFO - 18.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 18.self_attn.q_proj proxy error before VPTQ: 1.713288611426833e-06, 0.00010304831084795296, 0.01662607118487358
INFO - 18.self_attn.q_proj 1st error time: 2.299494743347168
INFO - 18.self_attn.q_proj proxy error after VPTQ: 1.1780539352912456e-06, 0.00010304831084795296, 0.011432054452598095
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3506.788818359375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 18.self_attn.q_proj residual time: 0.3172571659088135
INFO - 18.self_attn.q_proj 2ed gptq time: 2.590794086456299
INFO - 18.self_attn.q_proj 2ed error time: 0.0002865791320800781
INFO - 18.self_attn.q_proj proxy error after residual VPTQ: 4.1237404957428225e-07, 0.00010304831084795296, 0.004001754801720381
INFO - load Hessian from /workspace/hessians/quip-format/hessians/18_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:48:07 18.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:18.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 26 iterations, error 1917.20654296875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 18.self_attn.k_proj 1st kmeans time: 2.6386640071868896
INFO - 18.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 18.self_attn.k_proj proxy error before VPTQ: 3.677920119571354e-07, 2.1165786165511236e-05, 0.017376722767949104
INFO - 18.self_attn.k_proj 1st error time: 1.5496811866760254
INFO - 18.self_attn.k_proj proxy error after VPTQ: 2.990539371694467e-07, 2.1165786165511236e-05, 0.01412911992520094
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 830.2042846679688
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 18.self_attn.k_proj residual time: 0.09146451950073242
INFO - 18.self_attn.k_proj 2ed gptq time: 2.048344373703003
INFO - 18.self_attn.k_proj 2ed error time: 0.0003631114959716797
INFO - 18.self_attn.k_proj proxy error after residual VPTQ: 1.0115840609614679e-07, 2.1165786165511236e-05, 0.004779336042702198
INFO - load Hessian from /workspace/hessians/quip-format/hessians/18_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:48:15 18.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:18.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 25 iterations, error 2421.72802734375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 18.self_attn.v_proj 1st kmeans time: 2.5630483627319336
INFO - 18.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 18.self_attn.v_proj proxy error before VPTQ: 1.5219879401229264e-07, 5.336845788406208e-06, 0.028518492355942726
INFO - 18.self_attn.v_proj 1st error time: 1.5326406955718994
INFO - 18.self_attn.v_proj proxy error after VPTQ: 1.1978839609128045e-07, 5.336845788406208e-06, 0.022445542737841606
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1015.4124755859375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 18.self_attn.v_proj residual time: 0.09112548828125
INFO - 18.self_attn.v_proj 2ed gptq time: 2.0589680671691895
INFO - 18.self_attn.v_proj 2ed error time: 0.00038886070251464844
INFO - 18.self_attn.v_proj proxy error after residual VPTQ: 3.937272552434479e-08, 5.336845788406208e-06, 0.007377527188509703
INFO - load Hessian from /workspace/hessians/quip-format/hessians/18_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:48:23 18.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:18.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 62 iterations, error 17756.919921875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 18.self_attn.o_proj 1st kmeans time: 24.000107049942017
INFO - 18.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 18.self_attn.o_proj proxy error before VPTQ: 2.4060045689111575e-06, 0.0005243669147603214, 0.0045883990824222565
INFO - 18.self_attn.o_proj 1st error time: 2.3146770000457764
INFO - 18.self_attn.o_proj proxy error after VPTQ: 1.5062671536725247e-06, 0.0005243669147603214, 0.002872544340789318
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 6928.21533203125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 18.self_attn.o_proj residual time: 0.31476712226867676
INFO - 18.self_attn.o_proj 2ed gptq time: 2.5920915603637695
INFO - 18.self_attn.o_proj 2ed error time: 0.0009982585906982422
INFO - 18.self_attn.o_proj proxy error after residual VPTQ: 4.891420530839241e-07, 0.0005243669147603214, 0.0009328240412287414
INFO - load Hessian from /workspace/hessians/quip-format/hessians/18_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:48:54 18.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:18.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 49644.80859375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 18.mlp.gate_proj 1st kmeans time: 132.49943041801453
INFO - 18.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 18.mlp.gate_proj proxy error before VPTQ: 9.991732440539636e-06, 0.00034977399627678096, 0.028566252440214157
INFO - 18.mlp.gate_proj 1st error time: 4.752511024475098
INFO - 18.mlp.gate_proj proxy error after VPTQ: 7.721543624938931e-06, 0.00034977399627678096, 0.02207580767571926
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 99 iterations, error 18410.8046875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 18.mlp.gate_proj residual time: 1.048386812210083
INFO - 18.mlp.gate_proj 2ed gptq time: 5.123234510421753
INFO - 18.mlp.gate_proj 2ed error time: 0.0012497901916503906
INFO - 18.mlp.gate_proj proxy error after residual VPTQ: 2.568859144957969e-06, 0.00034977399627678096, 0.007344339974224567
INFO - load Hessian from /workspace/hessians/quip-format/hessians/18_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:51:20 18.mlp.up_proj
INFO - enabling norm dim 0, layer_name:18.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 66451.75
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 18.mlp.up_proj 1st kmeans time: 132.40602374076843
INFO - 18.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 18.mlp.up_proj proxy error before VPTQ: 7.318789812416071e-06, 0.00012087317736586556, 0.060549329966306686
INFO - 18.mlp.up_proj 1st error time: 4.770554065704346
INFO - 18.mlp.up_proj proxy error after VPTQ: 5.474021691043163e-06, 0.00012087317736586556, 0.04528731480240822
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 97 iterations, error 24052.38671875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 18.mlp.up_proj residual time: 1.117081880569458
INFO - 18.mlp.up_proj 2ed gptq time: 5.0449538230896
INFO - 18.mlp.up_proj 2ed error time: 0.01866316795349121
INFO - 18.mlp.up_proj proxy error after residual VPTQ: 1.781270270839741e-06, 0.00012087317736586556, 0.014736687764525414
INFO - load Hessian from /workspace/hessians/quip-format/hessians/18_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:53:48 18.mlp.down_proj
INFO - enabling norm dim 0, layer_name:18.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 84555.171875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 18.mlp.down_proj 1st kmeans time: 132.3681218624115
INFO - 18.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 18.mlp.down_proj proxy error before VPTQ: 6.39392965240404e-07, 1.4837990420346614e-05, 0.04309161379933357
INFO - 18.mlp.down_proj 1st error time: 8.52637004852295
INFO - 18.mlp.down_proj proxy error after VPTQ: 4.986270027984574e-07, 1.4837990420346614e-05, 0.03360475227236748
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 29794.3046875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 18.mlp.down_proj residual time: 1.0203864574432373
INFO - 18.mlp.down_proj 2ed gptq time: 9.608323097229004
INFO - 18.mlp.down_proj 2ed error time: 0.002341032028198242
INFO - 18.mlp.down_proj proxy error after residual VPTQ: 1.6218312737237284e-07, 1.4837990420346614e-05, 0.010930262506008148
INFO - ----Quantizing layer 19 ...---- 2026-02-23 10:56:23 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/19_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:56:24 19.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:19.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 87 iterations, error 9985.0673828125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 19.self_attn.q_proj 1st kmeans time: 33.2838077545166
INFO - 19.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 19.self_attn.q_proj proxy error before VPTQ: 1.550508613945567e-06, 7.275758252944797e-05, 0.021310612559318542
INFO - 19.self_attn.q_proj 1st error time: 2.3282241821289062
INFO - 19.self_attn.q_proj proxy error after VPTQ: 1.0519233910599723e-06, 7.275758252944797e-05, 0.014457921497523785
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 99 iterations, error 4052.29833984375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 19.self_attn.q_proj residual time: 0.3186044692993164
INFO - 19.self_attn.q_proj 2ed gptq time: 2.59800386428833
INFO - 19.self_attn.q_proj 2ed error time: 0.0006012916564941406
INFO - 19.self_attn.q_proj proxy error after residual VPTQ: 3.580847760531469e-07, 7.275758252944797e-05, 0.004921614658087492
INFO - load Hessian from /workspace/hessians/quip-format/hessians/19_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:57:04 19.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:19.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 20 iterations, error 2169.5400390625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 19.self_attn.k_proj 1st kmeans time: 2.0804567337036133
INFO - 19.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 19.self_attn.k_proj proxy error before VPTQ: 3.94809262616036e-07, 1.9813516701105982e-05, 0.01992625929415226
INFO - 19.self_attn.k_proj 1st error time: 1.5647170543670654
INFO - 19.self_attn.k_proj proxy error after VPTQ: 3.1077897233444673e-07, 1.9813516701105982e-05, 0.015685200691223145
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 929.8482666015625
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 19.self_attn.k_proj residual time: 0.09703731536865234
INFO - 19.self_attn.k_proj 2ed gptq time: 2.055288076400757
INFO - 19.self_attn.k_proj 2ed error time: 0.0005166530609130859
INFO - 19.self_attn.k_proj proxy error after residual VPTQ: 1.0416918883038306e-07, 1.9813516701105982e-05, 0.0052574812434613705
INFO - load Hessian from /workspace/hessians/quip-format/hessians/19_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:57:11 19.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:19.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 25 iterations, error 2719.802734375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 19.self_attn.v_proj 1st kmeans time: 2.560828924179077
INFO - 19.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 19.self_attn.v_proj proxy error before VPTQ: 1.754280134491637e-07, 5.329283794708317e-06, 0.03291774541139603
INFO - 19.self_attn.v_proj 1st error time: 1.5653982162475586
INFO - 19.self_attn.v_proj proxy error after VPTQ: 1.345224234228226e-07, 5.329283794708317e-06, 0.025242120027542114
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1100.966064453125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 19.self_attn.v_proj residual time: 0.10415410995483398
INFO - 19.self_attn.v_proj 2ed gptq time: 2.06789493560791
INFO - 19.self_attn.v_proj 2ed error time: 0.00047707557678222656
INFO - 19.self_attn.v_proj proxy error after residual VPTQ: 4.334289371854538e-08, 5.329283794708317e-06, 0.008132968097925186
INFO - load Hessian from /workspace/hessians/quip-format/hessians/19_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:57:19 19.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:19.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 66 iterations, error 19788.525390625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 19.self_attn.o_proj 1st kmeans time: 25.40964961051941
INFO - 19.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 19.self_attn.o_proj proxy error before VPTQ: 3.123667738691438e-06, 0.0023002028465270996, 0.001357996603474021
INFO - 19.self_attn.o_proj 1st error time: 2.336104393005371
INFO - 19.self_attn.o_proj proxy error after VPTQ: 2.2845899820822524e-06, 0.0023002028465270996, 0.0009932124521583319
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7638.8994140625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 19.self_attn.o_proj residual time: 0.31828904151916504
INFO - 19.self_attn.o_proj 2ed gptq time: 2.598517417907715
INFO - 19.self_attn.o_proj 2ed error time: 0.0011980533599853516
INFO - 19.self_attn.o_proj proxy error after residual VPTQ: 7.432331585732754e-07, 0.0023002028465270996, 0.00032311634276993573
INFO - load Hessian from /workspace/hessians/quip-format/hessians/19_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 10:57:51 19.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:19.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 47784.40234375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 19.mlp.gate_proj 1st kmeans time: 132.60345125198364
INFO - 19.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 19.mlp.gate_proj proxy error before VPTQ: 9.354991561849602e-06, 0.00029558525420725346, 0.03164904564619064
INFO - 19.mlp.gate_proj 1st error time: 4.863104581832886
INFO - 19.mlp.gate_proj proxy error after VPTQ: 6.999654488026863e-06, 0.00029558525420725346, 0.023680662736296654
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 84 iterations, error 17788.400390625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 19.mlp.gate_proj residual time: 0.9332802295684814
INFO - 19.mlp.gate_proj 2ed gptq time: 5.062713146209717
INFO - 19.mlp.gate_proj 2ed error time: 0.0014100074768066406
INFO - 19.mlp.gate_proj proxy error after residual VPTQ: 2.3183417852123966e-06, 0.00029558525420725346, 0.00784322526305914
INFO - load Hessian from /workspace/hessians/quip-format/hessians/19_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:00:17 19.mlp.up_proj
INFO - enabling norm dim 0, layer_name:19.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 61354.578125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 19.mlp.up_proj 1st kmeans time: 132.41494965553284
INFO - 19.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 19.mlp.up_proj proxy error before VPTQ: 7.00532655173447e-06, 0.0002185507328249514, 0.03205354884266853
INFO - 19.mlp.up_proj 1st error time: 4.833942651748657
INFO - 19.mlp.up_proj proxy error after VPTQ: 5.233355750533519e-06, 0.0002185507328249514, 0.023945724591612816
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 22303.23828125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 19.mlp.up_proj residual time: 1.0460009574890137
INFO - 19.mlp.up_proj 2ed gptq time: 5.149681091308594
INFO - 19.mlp.up_proj 2ed error time: 0.0014514923095703125
INFO - 19.mlp.up_proj proxy error after residual VPTQ: 1.6957587831711862e-06, 0.0002185507328249514, 0.007759108208119869
INFO - load Hessian from /workspace/hessians/quip-format/hessians/19_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:02:45 19.mlp.down_proj
INFO - enabling norm dim 0, layer_name:19.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 85905.8828125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 19.mlp.down_proj 1st kmeans time: 132.55455446243286
INFO - 19.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 19.mlp.down_proj proxy error before VPTQ: 6.399126846190484e-07, 1.5156271729210857e-05, 0.042220983654260635
INFO - 19.mlp.down_proj 1st error time: 8.65230679512024
INFO - 19.mlp.down_proj proxy error after VPTQ: 4.885904445472988e-07, 1.5156271729210857e-05, 0.03223684802651405
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 30301.240234375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 19.mlp.down_proj residual time: 1.0180723667144775
INFO - 19.mlp.down_proj 2ed gptq time: 9.622864246368408
INFO - 19.mlp.down_proj 2ed error time: 0.0025310516357421875
INFO - 19.mlp.down_proj proxy error after residual VPTQ: 1.5833789746011462e-07, 1.5156271729210857e-05, 0.010447021573781967
INFO - ----Quantizing layer 20 ...---- 2026-02-23 11:05:22 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/20_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:05:22 20.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:20.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 83 iterations, error 9221.466796875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 20.self_attn.q_proj 1st kmeans time: 31.60954213142395
INFO - 20.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 20.self_attn.q_proj proxy error before VPTQ: 1.5731571920696297e-06, 7.559877121821046e-05, 0.020809296518564224
INFO - 20.self_attn.q_proj 1st error time: 2.307929277420044
INFO - 20.self_attn.q_proj proxy error after VPTQ: 1.0946645261356025e-06, 7.559877121821046e-05, 0.01447992492467165
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3800.4951171875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 20.self_attn.q_proj residual time: 0.32019495964050293
INFO - 20.self_attn.q_proj 2ed gptq time: 2.6161677837371826
INFO - 20.self_attn.q_proj 2ed error time: 0.0005450248718261719
INFO - 20.self_attn.q_proj proxy error after residual VPTQ: 3.7776413819301524e-07, 7.559877121821046e-05, 0.004996961448341608
INFO - load Hessian from /workspace/hessians/quip-format/hessians/20_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:06:01 20.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:20.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 2248.327392578125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 20.self_attn.k_proj 1st kmeans time: 2.263059616088867
INFO - 20.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 20.self_attn.k_proj proxy error before VPTQ: 3.5763150663115084e-07, 2.1780091628897935e-05, 0.016420111060142517
INFO - 20.self_attn.k_proj 1st error time: 1.5373201370239258
INFO - 20.self_attn.k_proj proxy error after VPTQ: 2.786056825243577e-07, 2.1780091628897935e-05, 0.012791759334504604
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 947.8876953125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 20.self_attn.k_proj residual time: 0.09529709815979004
INFO - 20.self_attn.k_proj 2ed gptq time: 2.065343141555786
INFO - 20.self_attn.k_proj 2ed error time: 0.00044798851013183594
INFO - 20.self_attn.k_proj proxy error after residual VPTQ: 9.263438016660075e-08, 2.1780091628897935e-05, 0.004253167659044266
INFO - load Hessian from /workspace/hessians/quip-format/hessians/20_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:06:08 20.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:20.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 23 iterations, error 3544.58056640625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 20.self_attn.v_proj 1st kmeans time: 2.3741304874420166
INFO - 20.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 20.self_attn.v_proj proxy error before VPTQ: 2.55850125085999e-07, 1.1406516023271251e-05, 0.022430172190070152
INFO - 20.self_attn.v_proj 1st error time: 1.549309492111206
INFO - 20.self_attn.v_proj proxy error after VPTQ: 1.5228893346375116e-07, 1.1406516023271251e-05, 0.013351047411561012
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 77 iterations, error 1282.39794921875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 20.self_attn.v_proj residual time: 0.07879328727722168
INFO - 20.self_attn.v_proj 2ed gptq time: 2.0461246967315674
INFO - 20.self_attn.v_proj 2ed error time: 0.00045561790466308594
INFO - 20.self_attn.v_proj proxy error after residual VPTQ: 4.921168539340215e-08, 1.1406516023271251e-05, 0.004314348567277193
INFO - load Hessian from /workspace/hessians/quip-format/hessians/20_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:06:16 20.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:20.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 62 iterations, error 16057.2275390625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 20.self_attn.o_proj 1st kmeans time: 24.116698265075684
INFO - 20.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 20.self_attn.o_proj proxy error before VPTQ: 2.365567752349307e-06, 0.002592078410089016, 0.0009126142831519246
INFO - 20.self_attn.o_proj 1st error time: 2.361841917037964
INFO - 20.self_attn.o_proj proxy error after VPTQ: 1.806227714951092e-06, 0.002592078410089016, 0.0006968260277062654
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 6268.39306640625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 20.self_attn.o_proj residual time: 0.32045459747314453
INFO - 20.self_attn.o_proj 2ed gptq time: 2.5906646251678467
INFO - 20.self_attn.o_proj 2ed error time: 0.001405477523803711
INFO - 20.self_attn.o_proj proxy error after residual VPTQ: 6.005316208756994e-07, 0.002592078410089016, 0.00023167957260739058
INFO - load Hessian from /workspace/hessians/quip-format/hessians/20_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:06:47 20.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:20.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 56909.66796875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 20.mlp.gate_proj 1st kmeans time: 132.51074266433716
INFO - 20.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 20.mlp.gate_proj proxy error before VPTQ: 9.727318683872e-06, 0.00030400967807509005, 0.03199674189090729
INFO - 20.mlp.gate_proj 1st error time: 4.862478017807007
INFO - 20.mlp.gate_proj proxy error after VPTQ: 7.391001872747438e-06, 0.00030400967807509005, 0.02431173250079155
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 91 iterations, error 20731.49609375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 20.mlp.gate_proj residual time: 1.148667573928833
INFO - 20.mlp.gate_proj 2ed gptq time: 5.080219745635986
INFO - 20.mlp.gate_proj 2ed error time: 0.0014510154724121094
INFO - 20.mlp.gate_proj proxy error after residual VPTQ: 2.417422365397215e-06, 0.00030400967807509005, 0.007951794192194939
INFO - load Hessian from /workspace/hessians/quip-format/hessians/20_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:09:14 20.mlp.up_proj
INFO - enabling norm dim 0, layer_name:20.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 70319.8203125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 20.mlp.up_proj 1st kmeans time: 132.4185893535614
INFO - 20.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 20.mlp.up_proj proxy error before VPTQ: 7.152398666221416e-06, 0.00015261941007338464, 0.04686427861452103
INFO - 20.mlp.up_proj 1st error time: 4.757037401199341
INFO - 20.mlp.up_proj proxy error after VPTQ: 5.309960215527099e-06, 0.00015261941007338464, 0.03479216992855072
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 93 iterations, error 25249.26953125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 20.mlp.up_proj residual time: 1.0330584049224854
INFO - 20.mlp.up_proj 2ed gptq time: 5.07127571105957
INFO - 20.mlp.up_proj 2ed error time: 0.0011997222900390625
INFO - 20.mlp.up_proj proxy error after residual VPTQ: 1.7064182884496404e-06, 0.00015261941007338464, 0.011180873028934002
INFO - load Hessian from /workspace/hessians/quip-format/hessians/20_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:11:41 20.mlp.down_proj
INFO - enabling norm dim 0, layer_name:20.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 83902.6875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 20.mlp.down_proj 1st kmeans time: 132.5589120388031
INFO - 20.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 20.mlp.down_proj proxy error before VPTQ: 6.008799573464785e-07, 3.132262281724252e-05, 0.019183577969670296
INFO - 20.mlp.down_proj 1st error time: 9.333213567733765
INFO - 20.mlp.down_proj proxy error after VPTQ: 4.587591320159845e-07, 3.132262281724252e-05, 0.014646255411207676
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 99 iterations, error 29285.373046875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 20.mlp.down_proj residual time: 1.0172240734100342
INFO - 20.mlp.down_proj 2ed gptq time: 9.614398956298828
INFO - 20.mlp.down_proj 2ed error time: 0.0026273727416992188
INFO - 20.mlp.down_proj proxy error after residual VPTQ: 1.4799856273839396e-07, 3.132262281724252e-05, 0.004724973347038031
INFO - ----Quantizing layer 21 ...---- 2026-02-23 11:14:18 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/21_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:14:19 21.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:21.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 58 iterations, error 10413.3271484375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 21.self_attn.q_proj 1st kmeans time: 22.326014757156372
INFO - 21.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 21.self_attn.q_proj proxy error before VPTQ: 1.3962089724373072e-06, 6.836029206169769e-05, 0.020424269139766693
INFO - 21.self_attn.q_proj 1st error time: 2.391555070877075
INFO - 21.self_attn.q_proj proxy error after VPTQ: 9.035063612827798e-07, 6.836029206169769e-05, 0.013216829858720303
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 4166.30615234375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 21.self_attn.q_proj residual time: 0.31389784812927246
INFO - 21.self_attn.q_proj 2ed gptq time: 2.6027090549468994
INFO - 21.self_attn.q_proj 2ed error time: 0.0003685951232910156
INFO - 21.self_attn.q_proj proxy error after residual VPTQ: 3.054821888781589e-07, 6.836029206169769e-05, 0.004468708299100399
INFO - load Hessian from /workspace/hessians/quip-format/hessians/21_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:14:48 21.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:21.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 23 iterations, error 2104.08349609375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 21.self_attn.k_proj 1st kmeans time: 2.362994432449341
INFO - 21.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 21.self_attn.k_proj proxy error before VPTQ: 3.140934552448016e-07, 1.99962014448829e-05, 0.015707656741142273
INFO - 21.self_attn.k_proj 1st error time: 1.556903600692749
INFO - 21.self_attn.k_proj proxy error after VPTQ: 2.427276513117249e-07, 1.99962014448829e-05, 0.012138688005506992
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 880.5860595703125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 21.self_attn.k_proj residual time: 0.09120059013366699
INFO - 21.self_attn.k_proj 2ed gptq time: 2.0497286319732666
INFO - 21.self_attn.k_proj 2ed error time: 0.00039005279541015625
INFO - 21.self_attn.k_proj proxy error after residual VPTQ: 8.067050316640234e-08, 1.99962014448829e-05, 0.004034291487187147
INFO - load Hessian from /workspace/hessians/quip-format/hessians/21_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:14:56 21.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:21.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 2881.85498046875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 21.self_attn.v_proj 1st kmeans time: 2.267716646194458
INFO - 21.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 21.self_attn.v_proj proxy error before VPTQ: 2.2895591200722265e-07, 3.2096409995574504e-05, 0.007133380509912968
INFO - 21.self_attn.v_proj 1st error time: 1.5360236167907715
INFO - 21.self_attn.v_proj proxy error after VPTQ: 1.499989821240888e-07, 3.2096409995574504e-05, 0.004673388320952654
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1116.1285400390625
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 21.self_attn.v_proj residual time: 0.0914621353149414
INFO - 21.self_attn.v_proj 2ed gptq time: 2.0543212890625
INFO - 21.self_attn.v_proj 2ed error time: 0.0004987716674804688
INFO - 21.self_attn.v_proj proxy error after residual VPTQ: 4.871252912153068e-08, 3.2096409995574504e-05, 0.0015176939778029919
INFO - load Hessian from /workspace/hessians/quip-format/hessians/21_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:15:03 21.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:21.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 51 iterations, error 19552.900390625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 21.self_attn.o_proj 1st kmeans time: 19.73521637916565
INFO - 21.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 21.self_attn.o_proj proxy error before VPTQ: 2.2413116766983876e-06, 0.0006827758043073118, 0.0032826466485857964
INFO - 21.self_attn.o_proj 1st error time: 2.3117120265960693
INFO - 21.self_attn.o_proj proxy error after VPTQ: 1.5868698710619356e-06, 0.0006827758043073118, 0.002324144821614027
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7419.64501953125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 21.self_attn.o_proj residual time: 0.31476330757141113
INFO - 21.self_attn.o_proj 2ed gptq time: 2.598757266998291
INFO - 21.self_attn.o_proj 2ed error time: 0.0010166168212890625
INFO - 21.self_attn.o_proj proxy error after residual VPTQ: 5.066025892119796e-07, 0.0006827758043073118, 0.0007419750327244401
INFO - load Hessian from /workspace/hessians/quip-format/hessians/21_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:15:30 21.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:21.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 60403.921875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 21.mlp.gate_proj 1st kmeans time: 132.40507745742798
INFO - 21.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 21.mlp.gate_proj proxy error before VPTQ: 9.815380508371163e-06, 0.000433427921961993, 0.02264593541622162
INFO - 21.mlp.gate_proj 1st error time: 4.763755798339844
INFO - 21.mlp.gate_proj proxy error after VPTQ: 7.4289987423981074e-06, 0.000433427921961993, 0.01714010164141655
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 84 iterations, error 21814.41015625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 21.mlp.gate_proj residual time: 0.9304962158203125
INFO - 21.mlp.gate_proj 2ed gptq time: 5.157294273376465
INFO - 21.mlp.gate_proj 2ed error time: 0.001331329345703125
INFO - 21.mlp.gate_proj proxy error after residual VPTQ: 2.415074959571939e-06, 0.000433427921961993, 0.005572033580392599
INFO - load Hessian from /workspace/hessians/quip-format/hessians/21_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:17:55 21.mlp.up_proj
INFO - enabling norm dim 0, layer_name:21.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 72618.734375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 21.mlp.up_proj 1st kmeans time: 132.90535068511963
INFO - 21.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 21.mlp.up_proj proxy error before VPTQ: 7.160256245697383e-06, 0.00011976710084127262, 0.05978483334183693
INFO - 21.mlp.up_proj 1st error time: 4.74938178062439
INFO - 21.mlp.up_proj proxy error after VPTQ: 5.338490154827014e-06, 0.00011976710084127262, 0.04457392916083336
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 25825.19140625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 21.mlp.up_proj residual time: 1.0245063304901123
INFO - 21.mlp.up_proj 2ed gptq time: 5.07159423828125
INFO - 21.mlp.up_proj 2ed error time: 0.0012636184692382812
INFO - 21.mlp.up_proj proxy error after residual VPTQ: 1.7081094938475871e-06, 0.00011976710084127262, 0.014261925593018532
INFO - load Hessian from /workspace/hessians/quip-format/hessians/21_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:20:23 21.mlp.down_proj
INFO - enabling norm dim 0, layer_name:21.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 87634.9140625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 21.mlp.down_proj 1st kmeans time: 132.3727753162384
INFO - 21.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 21.mlp.down_proj proxy error before VPTQ: 6.054015670997615e-07, 2.3005739421932958e-05, 0.026315240189433098
INFO - 21.mlp.down_proj 1st error time: 8.820370435714722
INFO - 21.mlp.down_proj proxy error after VPTQ: 4.7492741828136786e-07, 2.3005739421932958e-05, 0.020643867552280426
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 30390.240234375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 21.mlp.down_proj residual time: 1.0286493301391602
INFO - 21.mlp.down_proj 2ed gptq time: 9.614780187606812
INFO - 21.mlp.down_proj 2ed error time: 0.002324342727661133
INFO - 21.mlp.down_proj proxy error after residual VPTQ: 1.5282324739018804e-07, 2.3005739421932958e-05, 0.006642831489443779
INFO - ----Quantizing layer 22 ...---- 2026-02-23 11:23:00 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/22_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:23:00 22.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:22.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 62 iterations, error 9705.3935546875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 22.self_attn.q_proj 1st kmeans time: 23.86088728904724
INFO - 22.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 22.self_attn.q_proj proxy error before VPTQ: 1.2452237569959834e-06, 6.647212285315618e-05, 0.01873302273452282
INFO - 22.self_attn.q_proj 1st error time: 2.4084928035736084
INFO - 22.self_attn.q_proj proxy error after VPTQ: 8.239877615778823e-07, 6.647212285315618e-05, 0.012395990081131458
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3890.540771484375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 22.self_attn.q_proj residual time: 0.35717129707336426
INFO - 22.self_attn.q_proj 2ed gptq time: 2.6015148162841797
INFO - 22.self_attn.q_proj 2ed error time: 0.0002956390380859375
INFO - 22.self_attn.q_proj proxy error after residual VPTQ: 2.8014386543873115e-07, 6.647212285315618e-05, 0.004214456304907799
INFO - load Hessian from /workspace/hessians/quip-format/hessians/22_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:23:31 22.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:22.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 26 iterations, error 1967.8994140625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 22.self_attn.k_proj 1st kmeans time: 2.6460416316986084
INFO - 22.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 22.self_attn.k_proj proxy error before VPTQ: 2.855546199498349e-07, 1.940340553119313e-05, 0.01471672672778368
INFO - 22.self_attn.k_proj 1st error time: 1.5441322326660156
INFO - 22.self_attn.k_proj proxy error after VPTQ: 2.213845391452196e-07, 1.940340553119313e-05, 0.011409571394324303
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 819.5472412109375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 22.self_attn.k_proj residual time: 0.09144473075866699
INFO - 22.self_attn.k_proj 2ed gptq time: 2.0555248260498047
INFO - 22.self_attn.k_proj 2ed error time: 0.0004980564117431641
INFO - 22.self_attn.k_proj proxy error after residual VPTQ: 7.408951319121115e-08, 1.940340553119313e-05, 0.003818376688286662
INFO - load Hessian from /workspace/hessians/quip-format/hessians/22_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:23:39 22.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:22.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 27 iterations, error 2821.35791015625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 22.self_attn.v_proj 1st kmeans time: 2.7432379722595215
INFO - 22.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 22.self_attn.v_proj proxy error before VPTQ: 2.305551305425979e-07, 1.1496453225845471e-05, 0.02005445770919323
INFO - 22.self_attn.v_proj 1st error time: 1.5600190162658691
INFO - 22.self_attn.v_proj proxy error after VPTQ: 1.43581189604447e-07, 1.1496453225845471e-05, 0.012489172630012035
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1037.5177001953125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 22.self_attn.v_proj residual time: 0.09388279914855957
INFO - 22.self_attn.v_proj 2ed gptq time: 2.055945873260498
INFO - 22.self_attn.v_proj 2ed error time: 0.0003867149353027344
INFO - 22.self_attn.v_proj proxy error after residual VPTQ: 4.626129168627813e-08, 1.1496453225845471e-05, 0.004023962188512087
INFO - load Hessian from /workspace/hessians/quip-format/hessians/22_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:23:47 22.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:22.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 61 iterations, error 18733.048828125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 22.self_attn.o_proj 1st kmeans time: 23.72086215019226
INFO - 22.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 22.self_attn.o_proj proxy error before VPTQ: 2.114783455908764e-06, 0.0005175011465325952, 0.004086528904736042
INFO - 22.self_attn.o_proj 1st error time: 2.306791305541992
INFO - 22.self_attn.o_proj proxy error after VPTQ: 1.3665978713106597e-06, 0.0005175011465325952, 0.0026407630648463964
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7133.41064453125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 22.self_attn.o_proj residual time: 0.31345272064208984
INFO - 22.self_attn.o_proj 2ed gptq time: 2.610337018966675
INFO - 22.self_attn.o_proj 2ed error time: 0.001064300537109375
INFO - 22.self_attn.o_proj proxy error after residual VPTQ: 4.388285219647514e-07, 0.0005175011465325952, 0.0008479759562760592
INFO - load Hessian from /workspace/hessians/quip-format/hessians/22_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:24:17 22.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:22.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 64056.5
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 22.mlp.gate_proj 1st kmeans time: 132.41111135482788
INFO - 22.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 22.mlp.gate_proj proxy error before VPTQ: 9.432041224499699e-06, 0.00033109867945313454, 0.028487099334597588
INFO - 22.mlp.gate_proj 1st error time: 4.790297508239746
INFO - 22.mlp.gate_proj proxy error after VPTQ: 7.2339553298661485e-06, 0.00033109867945313454, 0.02184833586215973
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 23015.041015625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 22.mlp.gate_proj residual time: 1.025254249572754
INFO - 22.mlp.gate_proj 2ed gptq time: 5.079181909561157
INFO - 22.mlp.gate_proj 2ed error time: 0.0012521743774414062
INFO - 22.mlp.gate_proj proxy error after residual VPTQ: 2.347894906051806e-06, 0.00033109867945313454, 0.007091224193572998
INFO - load Hessian from /workspace/hessians/quip-format/hessians/22_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:26:43 22.mlp.up_proj
INFO - enabling norm dim 0, layer_name:22.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 73812.4765625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 22.mlp.up_proj 1st kmeans time: 132.49991273880005
INFO - 22.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 22.mlp.up_proj proxy error before VPTQ: 7.200746040325612e-06, 0.00010412019037175924, 0.06915801763534546
INFO - 22.mlp.up_proj 1st error time: 4.836204767227173
INFO - 22.mlp.up_proj proxy error after VPTQ: 5.384954420151189e-06, 0.00010412019037175924, 0.051718637347221375
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 26109.8203125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 22.mlp.up_proj residual time: 1.0451538562774658
INFO - 22.mlp.up_proj 2ed gptq time: 5.060930252075195
INFO - 22.mlp.up_proj 2ed error time: 0.001135110855102539
INFO - 22.mlp.up_proj proxy error after residual VPTQ: 1.7225179362867493e-06, 0.00010412019037175924, 0.0165435541421175
INFO - load Hessian from /workspace/hessians/quip-format/hessians/22_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:29:11 22.mlp.down_proj
INFO - enabling norm dim 0, layer_name:22.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 91233.0078125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 22.mlp.down_proj 1st kmeans time: 132.36933732032776
INFO - 22.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 22.mlp.down_proj proxy error before VPTQ: 6.192460091369867e-07, 1.3004947504668962e-05, 0.04761618748307228
INFO - 22.mlp.down_proj 1st error time: 8.826003313064575
INFO - 22.mlp.down_proj proxy error after VPTQ: 5.069824169368076e-07, 1.3004947504668962e-05, 0.03898381069302559
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 31349.5234375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 22.mlp.down_proj residual time: 1.0354187488555908
INFO - 22.mlp.down_proj 2ed gptq time: 9.636527061462402
INFO - 22.mlp.down_proj 2ed error time: 0.0025136470794677734
INFO - 22.mlp.down_proj proxy error after residual VPTQ: 1.6336652208792657e-07, 1.3004947504668962e-05, 0.012561874464154243
INFO - ----Quantizing layer 23 ...---- 2026-02-23 11:31:47 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/23_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:31:48 23.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:23.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 74 iterations, error 10582.95703125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 23.self_attn.q_proj 1st kmeans time: 28.27237105369568
INFO - 23.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 23.self_attn.q_proj proxy error before VPTQ: 1.3564979326474713e-06, 0.00012849990162067115, 0.010556411929428577
INFO - 23.self_attn.q_proj 1st error time: 2.355757236480713
INFO - 23.self_attn.q_proj proxy error after VPTQ: 8.622374707556446e-07, 0.00012849990162067115, 0.0067100245505571365
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 4197.15869140625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 23.self_attn.q_proj residual time: 0.3176155090332031
INFO - 23.self_attn.q_proj 2ed gptq time: 2.5944817066192627
INFO - 23.self_attn.q_proj 2ed error time: 0.0003597736358642578
INFO - 23.self_attn.q_proj proxy error after residual VPTQ: 2.9139334856154164e-07, 0.00012849990162067115, 0.002267654286697507
INFO - load Hessian from /workspace/hessians/quip-format/hessians/23_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:32:23 23.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:23.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 25 iterations, error 2062.1396484375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 23.self_attn.k_proj 1st kmeans time: 2.555122137069702
INFO - 23.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 23.self_attn.k_proj proxy error before VPTQ: 3.0969397357694106e-07, 2.017758015426807e-05, 0.015348419547080994
INFO - 23.self_attn.k_proj 1st error time: 1.549696922302246
INFO - 23.self_attn.k_proj proxy error after VPTQ: 2.429882499654923e-07, 2.017758015426807e-05, 0.012042487040162086
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 845.2543334960938
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 23.self_attn.k_proj residual time: 0.09735298156738281
INFO - 23.self_attn.k_proj 2ed gptq time: 2.0420126914978027
INFO - 23.self_attn.k_proj 2ed error time: 0.0004570484161376953
INFO - 23.self_attn.k_proj proxy error after residual VPTQ: 8.049811839327958e-08, 2.017758015426807e-05, 0.00398948322981596
INFO - load Hessian from /workspace/hessians/quip-format/hessians/23_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:32:31 23.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:23.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 21 iterations, error 2336.35986328125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 23.self_attn.v_proj 1st kmeans time: 2.1864726543426514
INFO - 23.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 23.self_attn.v_proj proxy error before VPTQ: 1.8461314255091565e-07, 6.797944024583558e-06, 0.02715720236301422
INFO - 23.self_attn.v_proj 1st error time: 1.5602943897247314
INFO - 23.self_attn.v_proj proxy error after VPTQ: 1.4624004052166129e-07, 6.797944024583558e-06, 0.021512392908334732
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 929.9774780273438
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 23.self_attn.v_proj residual time: 0.09253120422363281
INFO - 23.self_attn.v_proj 2ed gptq time: 2.0466513633728027
INFO - 23.self_attn.v_proj 2ed error time: 0.000438690185546875
INFO - 23.self_attn.v_proj proxy error after residual VPTQ: 4.729214708731888e-08, 6.797944024583558e-06, 0.006956831086426973
INFO - load Hessian from /workspace/hessians/quip-format/hessians/23_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:32:38 23.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:23.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 68 iterations, error 20428.005859375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 23.self_attn.o_proj 1st kmeans time: 26.25889492034912
INFO - 23.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 23.self_attn.o_proj proxy error before VPTQ: 2.5671702132967766e-06, 0.0008222117321565747, 0.0031222738325595856
INFO - 23.self_attn.o_proj 1st error time: 2.318891763687134
INFO - 23.self_attn.o_proj proxy error after VPTQ: 1.6622938119326136e-06, 0.0008222117321565747, 0.002021734369918704
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7762.583984375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 23.self_attn.o_proj residual time: 0.31334424018859863
INFO - 23.self_attn.o_proj 2ed gptq time: 2.64870023727417
INFO - 23.self_attn.o_proj 2ed error time: 0.0010938644409179688
INFO - 23.self_attn.o_proj proxy error after residual VPTQ: 5.333698709364398e-07, 0.0008222117321565747, 0.0006487013888545334
INFO - load Hessian from /workspace/hessians/quip-format/hessians/23_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:33:12 23.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:23.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 66844.7109375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 23.mlp.gate_proj 1st kmeans time: 132.50226616859436
INFO - 23.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 23.mlp.gate_proj proxy error before VPTQ: 8.751317182031926e-06, 0.00035821314668282866, 0.024430474266409874
INFO - 23.mlp.gate_proj 1st error time: 4.775818109512329
INFO - 23.mlp.gate_proj proxy error after VPTQ: 6.699577170365956e-06, 0.00035821314668282866, 0.01870276778936386
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 94 iterations, error 23920.947265625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 23.mlp.gate_proj residual time: 1.0331871509552002
INFO - 23.mlp.gate_proj 2ed gptq time: 5.157143831253052
INFO - 23.mlp.gate_proj 2ed error time: 0.0014417171478271484
INFO - 23.mlp.gate_proj proxy error after residual VPTQ: 2.1580281099886633e-06, 0.00035821314668282866, 0.006024424452334642
INFO - load Hessian from /workspace/hessians/quip-format/hessians/23_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:35:38 23.mlp.up_proj
INFO - enabling norm dim 0, layer_name:23.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 73572.296875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 23.mlp.up_proj 1st kmeans time: 132.40099906921387
INFO - 23.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 23.mlp.up_proj proxy error before VPTQ: 7.101017217792105e-06, 0.00010164199920836836, 0.06986302137374878
INFO - 23.mlp.up_proj 1st error time: 4.810361862182617
INFO - 23.mlp.up_proj proxy error after VPTQ: 5.361119292501826e-06, 0.00010164199920836836, 0.052745118737220764
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 25954.130859375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 23.mlp.up_proj residual time: 1.0764234066009521
INFO - 23.mlp.up_proj 2ed gptq time: 5.06467342376709
INFO - 23.mlp.up_proj 2ed error time: 0.0017077922821044922
INFO - 23.mlp.up_proj proxy error after residual VPTQ: 1.714235281724541e-06, 0.00010164199920836836, 0.016865422949194908
INFO - load Hessian from /workspace/hessians/quip-format/hessians/23_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:38:05 23.mlp.down_proj
INFO - enabling norm dim 0, layer_name:23.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 92078.546875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 23.mlp.down_proj 1st kmeans time: 132.3735740184784
INFO - 23.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 23.mlp.down_proj proxy error before VPTQ: 6.079403078729229e-07, 9.353901987196878e-06, 0.0649932324886322
INFO - 23.mlp.down_proj 1st error time: 8.887686491012573
INFO - 23.mlp.down_proj proxy error after VPTQ: 5.085045700070623e-07, 9.353901987196878e-06, 0.05436282977461815
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 31493.673828125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 23.mlp.down_proj residual time: 1.0712072849273682
INFO - 23.mlp.down_proj 2ed gptq time: 9.628618717193604
INFO - 23.mlp.down_proj 2ed error time: 0.0025217533111572266
INFO - 23.mlp.down_proj proxy error after residual VPTQ: 1.635781075037812e-07, 9.353901987196878e-06, 0.01748768612742424
INFO - ----Quantizing layer 24 ...---- 2026-02-23 11:40:42 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/24_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:40:43 24.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:24.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 75 iterations, error 10499.28515625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 24.self_attn.q_proj 1st kmeans time: 28.634004831314087
INFO - 24.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 24.self_attn.q_proj proxy error before VPTQ: 1.237319338542875e-06, 7.940572686493397e-05, 0.015582242980599403
INFO - 24.self_attn.q_proj 1st error time: 2.328223943710327
INFO - 24.self_attn.q_proj proxy error after VPTQ: 8.449197821391863e-07, 7.940572686493397e-05, 0.010640539228916168
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 4142.001953125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 24.self_attn.q_proj residual time: 0.31313347816467285
INFO - 24.self_attn.q_proj 2ed gptq time: 2.6816911697387695
INFO - 24.self_attn.q_proj 2ed error time: 0.00032258033752441406
INFO - 24.self_attn.q_proj proxy error after residual VPTQ: 2.8463520607147075e-07, 7.940572686493397e-05, 0.003584567690268159
INFO - load Hessian from /workspace/hessians/quip-format/hessians/24_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:41:18 24.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:24.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 23 iterations, error 2143.460693359375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 24.self_attn.k_proj 1st kmeans time: 2.360710859298706
INFO - 24.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 24.self_attn.k_proj proxy error before VPTQ: 2.815469599681819e-07, 1.8391892808722332e-05, 0.01530820969492197
INFO - 24.self_attn.k_proj 1st error time: 1.548388957977295
INFO - 24.self_attn.k_proj proxy error after VPTQ: 2.239440135554105e-07, 1.8391892808722332e-05, 0.01217623520642519
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 867.364501953125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 24.self_attn.k_proj residual time: 0.19128012657165527
INFO - 24.self_attn.k_proj 2ed gptq time: 2.074383020401001
INFO - 24.self_attn.k_proj 2ed error time: 0.0006115436553955078
INFO - 24.self_attn.k_proj proxy error after residual VPTQ: 7.401852286648136e-08, 1.8391892808722332e-05, 0.004024519119411707
INFO - load Hessian from /workspace/hessians/quip-format/hessians/24_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:41:26 24.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:24.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 33 iterations, error 2668.891357421875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 24.self_attn.v_proj 1st kmeans time: 3.2969722747802734
INFO - 24.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 24.self_attn.v_proj proxy error before VPTQ: 2.2310857161755848e-07, 1.3232602213975042e-05, 0.016860520467162132
INFO - 24.self_attn.v_proj 1st error time: 1.550722599029541
INFO - 24.self_attn.v_proj proxy error after VPTQ: 1.6134664804212662e-07, 1.3232602213975042e-05, 0.012193115428090096
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1009.131591796875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 24.self_attn.v_proj residual time: 0.10236692428588867
INFO - 24.self_attn.v_proj 2ed gptq time: 2.0863938331604004
INFO - 24.self_attn.v_proj 2ed error time: 0.0003752708435058594
INFO - 24.self_attn.v_proj proxy error after residual VPTQ: 5.203706621159654e-08, 1.3232602213975042e-05, 0.0039324890822172165
INFO - load Hessian from /workspace/hessians/quip-format/hessians/24_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:41:34 24.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:24.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 60 iterations, error 22245.197265625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 24.self_attn.o_proj 1st kmeans time: 23.229047775268555
INFO - 24.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 24.self_attn.o_proj proxy error before VPTQ: 2.031692929449491e-06, 4.887701652478427e-05, 0.041567448526620865
INFO - 24.self_attn.o_proj 1st error time: 2.318023920059204
INFO - 24.self_attn.o_proj proxy error after VPTQ: 1.3854983080818783e-06, 4.887701652478427e-05, 0.028346622362732887
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 8276.654296875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 24.self_attn.o_proj residual time: 0.31409573554992676
INFO - 24.self_attn.o_proj 2ed gptq time: 2.6028852462768555
INFO - 24.self_attn.o_proj 2ed error time: 0.0010995864868164062
INFO - 24.self_attn.o_proj proxy error after residual VPTQ: 4.3731660070989165e-07, 4.887701652478427e-05, 0.008947284892201424
INFO - load Hessian from /workspace/hessians/quip-format/hessians/24_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:42:05 24.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:24.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 64342.4609375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 24.mlp.gate_proj 1st kmeans time: 132.49971270561218
INFO - 24.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 24.mlp.gate_proj proxy error before VPTQ: 8.452392648905516e-06, 0.00014982811990194023, 0.0564139261841774
INFO - 24.mlp.gate_proj 1st error time: 4.765437364578247
INFO - 24.mlp.gate_proj proxy error after VPTQ: 6.526034212583909e-06, 0.00014982811990194023, 0.043556805700063705
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 87 iterations, error 23054.67578125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 24.mlp.gate_proj residual time: 0.9265480041503906
INFO - 24.mlp.gate_proj 2ed gptq time: 5.058645009994507
INFO - 24.mlp.gate_proj 2ed error time: 0.0013310909271240234
INFO - 24.mlp.gate_proj proxy error after residual VPTQ: 2.1177015696594026e-06, 0.00014982811990194023, 0.014134206809103489
INFO - load Hessian from /workspace/hessians/quip-format/hessians/24_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:44:31 24.mlp.up_proj
INFO - enabling norm dim 0, layer_name:24.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 71457.5625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 24.mlp.up_proj 1st kmeans time: 132.4045889377594
INFO - 24.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 24.mlp.up_proj proxy error before VPTQ: 6.909638159413589e-06, 0.00010626988660078496, 0.06501971930265427
INFO - 24.mlp.up_proj 1st error time: 4.845832586288452
INFO - 24.mlp.up_proj proxy error after VPTQ: 5.316815986589063e-06, 0.00010626988660078496, 0.050031255930662155
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 25255.30859375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 24.mlp.up_proj residual time: 1.0527818202972412
INFO - 24.mlp.up_proj 2ed gptq time: 5.058370590209961
INFO - 24.mlp.up_proj 2ed error time: 0.001173257827758789
INFO - 24.mlp.up_proj proxy error after residual VPTQ: 1.7052379917004146e-06, 0.00010626988660078496, 0.016046294942498207
INFO - load Hessian from /workspace/hessians/quip-format/hessians/24_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:46:58 24.mlp.down_proj
INFO - enabling norm dim 0, layer_name:24.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 5150.6318359375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 24.mlp.down_proj 1st kmeans time: 132.3677260875702
INFO - 24.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 24.mlp.down_proj proxy error before VPTQ: 5.8374858724619116e-08, 0.000378209981136024, 0.00015434510714840144
INFO - 24.mlp.down_proj 1st error time: 9.096100330352783
INFO - 24.mlp.down_proj proxy error after VPTQ: 4.7926352664262595e-08, 0.000378209981136024, 0.00012671889271587133
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1926.8876953125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 24.mlp.down_proj residual time: 1.0491585731506348
INFO - 24.mlp.down_proj 2ed gptq time: 10.191572904586792
INFO - 24.mlp.down_proj 2ed error time: 0.002866506576538086
INFO - 24.mlp.down_proj proxy error after residual VPTQ: 1.5870591951738788e-08, 0.000378209981136024, 4.196238296572119e-05
INFO - ----Quantizing layer 25 ...---- 2026-02-23 11:49:35 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/25_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:49:36 25.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:25.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 78 iterations, error 8314.1220703125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 25.self_attn.q_proj 1st kmeans time: 29.857226610183716
INFO - 25.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 25.self_attn.q_proj proxy error before VPTQ: 1.5179314232227625e-06, 0.00012955398415215313, 0.011716593988239765
INFO - 25.self_attn.q_proj 1st error time: 2.3102760314941406
INFO - 25.self_attn.q_proj proxy error after VPTQ: 9.644518286222592e-07, 0.00012955398415215313, 0.007444401271641254
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 97 iterations, error 3432.910400390625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 25.self_attn.q_proj residual time: 0.31139111518859863
INFO - 25.self_attn.q_proj 2ed gptq time: 2.7184157371520996
INFO - 25.self_attn.q_proj 2ed error time: 0.0005133152008056641
INFO - 25.self_attn.q_proj proxy error after residual VPTQ: 3.387228844076162e-07, 0.00012955398415215313, 0.002614530734717846
INFO - load Hessian from /workspace/hessians/quip-format/hessians/25_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:50:13 25.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:25.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 30 iterations, error 1552.335693359375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 25.self_attn.k_proj 1st kmeans time: 3.022303581237793
INFO - 25.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 25.self_attn.k_proj proxy error before VPTQ: 3.646370601018134e-07, 3.34689138981048e-05, 0.01089479774236679
INFO - 25.self_attn.k_proj 1st error time: 1.6465256214141846
INFO - 25.self_attn.k_proj proxy error after VPTQ: 2.897360218412359e-07, 3.34689138981048e-05, 0.008656869642436504
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 664.964599609375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 25.self_attn.k_proj residual time: 0.09847545623779297
INFO - 25.self_attn.k_proj 2ed gptq time: 2.124948024749756
INFO - 25.self_attn.k_proj 2ed error time: 0.0006308555603027344
INFO - 25.self_attn.k_proj proxy error after residual VPTQ: 9.924887933721038e-08, 3.34689138981048e-05, 0.002965404884889722
INFO - load Hessian from /workspace/hessians/quip-format/hessians/25_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:50:21 25.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:25.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 2313.7646484375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 25.self_attn.v_proj 1st kmeans time: 2.281132221221924
INFO - 25.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 25.self_attn.v_proj proxy error before VPTQ: 1.4745748444511264e-07, 5.449187483463902e-06, 0.027060452848672867
INFO - 25.self_attn.v_proj 1st error time: 1.548461675643921
INFO - 25.self_attn.v_proj proxy error after VPTQ: 1.0834938279913331e-07, 5.449187483463902e-06, 0.01988358423113823
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 977.5128173828125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 25.self_attn.v_proj residual time: 0.09138059616088867
INFO - 25.self_attn.v_proj 2ed gptq time: 2.0570662021636963
INFO - 25.self_attn.v_proj 2ed error time: 0.0005247592926025391
INFO - 25.self_attn.v_proj proxy error after residual VPTQ: 3.558388073088281e-08, 5.449187483463902e-06, 0.006530126091092825
INFO - load Hessian from /workspace/hessians/quip-format/hessians/25_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:50:29 25.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:25.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 86 iterations, error 17325.9921875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 25.self_attn.o_proj 1st kmeans time: 32.751150369644165
INFO - 25.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 25.self_attn.o_proj proxy error before VPTQ: 2.60561546383542e-06, 0.0010623130947351456, 0.0024527753703296185
INFO - 25.self_attn.o_proj 1st error time: 2.4121460914611816
INFO - 25.self_attn.o_proj proxy error after VPTQ: 2.1732819277531235e-06, 0.0010623130947351456, 0.0020458016078919172
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 6722.134765625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 25.self_attn.o_proj residual time: 0.3245987892150879
INFO - 25.self_attn.o_proj 2ed gptq time: 2.6184749603271484
INFO - 25.self_attn.o_proj 2ed error time: 0.0012974739074707031
INFO - 25.self_attn.o_proj proxy error after residual VPTQ: 7.271011099874158e-07, 0.0010623130947351456, 0.0006844508461654186
INFO - load Hessian from /workspace/hessians/quip-format/hessians/25_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:51:08 25.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:25.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 69883.0859375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 25.mlp.gate_proj 1st kmeans time: 132.5134744644165
INFO - 25.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 25.mlp.gate_proj proxy error before VPTQ: 1.0491016837477218e-05, 0.00022727373288944364, 0.046160269528627396
INFO - 25.mlp.gate_proj 1st error time: 4.765886545181274
INFO - 25.mlp.gate_proj proxy error after VPTQ: 8.20085551822558e-06, 0.00022727373288944364, 0.03608360514044762
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 92 iterations, error 25058.484375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 25.mlp.gate_proj residual time: 0.9526901245117188
INFO - 25.mlp.gate_proj 2ed gptq time: 5.05936598777771
INFO - 25.mlp.gate_proj 2ed error time: 0.0013093948364257812
INFO - 25.mlp.gate_proj proxy error after residual VPTQ: 2.645576842041919e-06, 0.00022727373288944364, 0.011640486307442188
INFO - load Hessian from /workspace/hessians/quip-format/hessians/25_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:53:34 25.mlp.up_proj
INFO - enabling norm dim 0, layer_name:25.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 70093.7265625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 25.mlp.up_proj 1st kmeans time: 132.36566734313965
INFO - 25.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 25.mlp.up_proj proxy error before VPTQ: 9.645236787036993e-06, 0.00015396227536257356, 0.06264675408601761
INFO - 25.mlp.up_proj 1st error time: 4.7512757778167725
INFO - 25.mlp.up_proj proxy error after VPTQ: 7.5606731115840375e-06, 0.00015396227536257356, 0.04910730943083763
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 98 iterations, error 25068.912109375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 25.mlp.up_proj residual time: 1.0373356342315674
INFO - 25.mlp.up_proj 2ed gptq time: 5.07179069519043
INFO - 25.mlp.up_proj 2ed error time: 0.0012676715850830078
INFO - 25.mlp.up_proj proxy error after residual VPTQ: 2.4421165107924026e-06, 0.00015396227536257356, 0.015861785039305687
INFO - load Hessian from /workspace/hessians/quip-format/hessians/25_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:56:01 25.mlp.down_proj
INFO - enabling norm dim 0, layer_name:25.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 87279.6015625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 25.mlp.down_proj 1st kmeans time: 132.39281296730042
INFO - 25.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 25.mlp.down_proj proxy error before VPTQ: 8.321081850226619e-07, 2.4783144908724353e-05, 0.03357556834816933
INFO - 25.mlp.down_proj 1st error time: 8.642708778381348
INFO - 25.mlp.down_proj proxy error after VPTQ: 6.535953502861958e-07, 2.4783144908724353e-05, 0.026372574269771576
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 30774.28125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 25.mlp.down_proj residual time: 1.0681946277618408
INFO - 25.mlp.down_proj 2ed gptq time: 9.645766973495483
INFO - 25.mlp.down_proj 2ed error time: 0.002525806427001953
INFO - 25.mlp.down_proj proxy error after residual VPTQ: 2.10075782547392e-07, 2.4783144908724353e-05, 0.008476559072732925
INFO - ----Quantizing layer 26 ...---- 2026-02-23 11:58:38 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/26_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:58:38 26.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:26.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 65 iterations, error 8745.7841796875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 26.self_attn.q_proj 1st kmeans time: 24.988290786743164
INFO - 26.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 26.self_attn.q_proj proxy error before VPTQ: 1.6409132967964979e-06, 0.00011645849008345976, 0.014090113341808319
INFO - 26.self_attn.q_proj 1st error time: 2.324498414993286
INFO - 26.self_attn.q_proj proxy error after VPTQ: 1.0855479786187061e-06, 0.00011645849008345976, 0.009321330115199089
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3584.708740234375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 26.self_attn.q_proj residual time: 0.3142411708831787
INFO - 26.self_attn.q_proj 2ed gptq time: 2.5942234992980957
INFO - 26.self_attn.q_proj 2ed error time: 0.0003123283386230469
INFO - 26.self_attn.q_proj proxy error after residual VPTQ: 3.732684490387328e-07, 0.00011645849008345976, 0.003205162938684225
INFO - load Hessian from /workspace/hessians/quip-format/hessians/26_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:59:10 26.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:26.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 25 iterations, error 1818.007080078125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 26.self_attn.k_proj 1st kmeans time: 2.5503227710723877
INFO - 26.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 26.self_attn.k_proj proxy error before VPTQ: 4.207197434880072e-07, 3.918178117601201e-05, 0.010737637057900429
INFO - 26.self_attn.k_proj 1st error time: 1.5361740589141846
INFO - 26.self_attn.k_proj proxy error after VPTQ: 3.2571600172559556e-07, 3.918178117601201e-05, 0.008312945254147053
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 775.0433959960938
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 26.self_attn.k_proj residual time: 0.0951986312866211
INFO - 26.self_attn.k_proj 2ed gptq time: 2.0573575496673584
INFO - 26.self_attn.k_proj 2ed error time: 0.0004093647003173828
INFO - 26.self_attn.k_proj proxy error after residual VPTQ: 1.09783691470966e-07, 3.918178117601201e-05, 0.0028019065503031015
INFO - load Hessian from /workspace/hessians/quip-format/hessians/26_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:59:18 26.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:26.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 2412.837890625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 26.self_attn.v_proj 1st kmeans time: 2.275808334350586
INFO - 26.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 26.self_attn.v_proj proxy error before VPTQ: 1.600139682977897e-07, 5.854102710145526e-06, 0.027333645150065422
INFO - 26.self_attn.v_proj 1st error time: 1.5777015686035156
INFO - 26.self_attn.v_proj proxy error after VPTQ: 1.2068757371253014e-07, 5.854102710145526e-06, 0.020615896210074425
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1012.74169921875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 26.self_attn.v_proj residual time: 0.09609341621398926
INFO - 26.self_attn.v_proj 2ed gptq time: 2.0779006481170654
INFO - 26.self_attn.v_proj 2ed error time: 0.00041866302490234375
INFO - 26.self_attn.v_proj proxy error after residual VPTQ: 3.96203638786119e-08, 5.854102710145526e-06, 0.006767965387552977
INFO - load Hessian from /workspace/hessians/quip-format/hessians/26_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 11:59:25 26.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:26.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 75 iterations, error 17481.68359375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 26.self_attn.o_proj 1st kmeans time: 28.65672469139099
INFO - 26.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 26.self_attn.o_proj proxy error before VPTQ: 1.683486971160164e-06, 5.74367368244566e-05, 0.029310282319784164
INFO - 26.self_attn.o_proj 1st error time: 2.311903715133667
INFO - 26.self_attn.o_proj proxy error after VPTQ: 1.0206940714851953e-06, 5.74367368244566e-05, 0.01777075231075287
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 6796.3427734375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 26.self_attn.o_proj residual time: 0.31087446212768555
INFO - 26.self_attn.o_proj 2ed gptq time: 2.588087320327759
INFO - 26.self_attn.o_proj 2ed error time: 0.001065969467163086
INFO - 26.self_attn.o_proj proxy error after residual VPTQ: 3.2977669661704567e-07, 5.74367368244566e-05, 0.005741564091295004
INFO - load Hessian from /workspace/hessians/quip-format/hessians/26_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:00:01 26.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:26.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 68123.9296875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 26.mlp.gate_proj 1st kmeans time: 132.49701285362244
INFO - 26.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 26.mlp.gate_proj proxy error before VPTQ: 1.0023108188761398e-05, 0.00017592226504348218, 0.05697464197874069
INFO - 26.mlp.gate_proj 1st error time: 4.884860992431641
INFO - 26.mlp.gate_proj proxy error after VPTQ: 7.850695510569494e-06, 0.00017592226504348218, 0.04462593421339989
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 24479.099609375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 26.mlp.gate_proj residual time: 1.0195441246032715
INFO - 26.mlp.gate_proj 2ed gptq time: 5.070606231689453
INFO - 26.mlp.gate_proj 2ed error time: 0.0014603137969970703
INFO - 26.mlp.gate_proj proxy error after residual VPTQ: 2.5403628569620196e-06, 0.00017592226504348218, 0.014440258033573627
INFO - load Hessian from /workspace/hessians/quip-format/hessians/26_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:02:27 26.mlp.up_proj
INFO - enabling norm dim 0, layer_name:26.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 68359.8671875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 26.mlp.up_proj 1st kmeans time: 132.40717196464539
INFO - 26.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 26.mlp.up_proj proxy error before VPTQ: 9.226499059877824e-06, 0.00014568015467375517, 0.06333394348621368
INFO - 26.mlp.up_proj 1st error time: 4.729902267456055
INFO - 26.mlp.up_proj proxy error after VPTQ: 7.231423296616413e-06, 0.00014568015467375517, 0.04963904246687889
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 100 iterations, error 24513.1796875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 26.mlp.up_proj residual time: 1.057356834411621
INFO - 26.mlp.up_proj 2ed gptq time: 5.074174642562866
INFO - 26.mlp.up_proj 2ed error time: 0.015331268310546875
INFO - 26.mlp.up_proj proxy error after residual VPTQ: 2.342366769880755e-06, 0.00014568015467375517, 0.016078831627964973
INFO - load Hessian from /workspace/hessians/quip-format/hessians/26_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:04:54 26.mlp.down_proj
INFO - enabling norm dim 0, layer_name:26.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 87680.1796875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 26.mlp.down_proj 1st kmeans time: 132.4111886024475
INFO - 26.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 26.mlp.down_proj proxy error before VPTQ: 8.310353791785019e-07, 3.1187242711894214e-05, 0.026646645739674568
INFO - 26.mlp.down_proj 1st error time: 8.760598421096802
INFO - 26.mlp.down_proj proxy error after VPTQ: 6.842209359092521e-07, 3.1187242711894214e-05, 0.021939128637313843
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 95 iterations, error 30569.6015625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 26.mlp.down_proj residual time: 0.965752124786377
INFO - 26.mlp.down_proj 2ed gptq time: 9.59741759300232
INFO - 26.mlp.down_proj 2ed error time: 0.0023949146270751953
INFO - 26.mlp.down_proj proxy error after residual VPTQ: 2.20825711494399e-07, 3.1187242711894214e-05, 0.007080642506480217
INFO - ----Quantizing layer 27 ...---- 2026-02-23 12:07:31 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/27_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:07:31 27.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:27.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 68 iterations, error 8531.9033203125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 27.self_attn.q_proj 1st kmeans time: 26.04008650779724
INFO - 27.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 27.self_attn.q_proj proxy error before VPTQ: 1.8551104403741192e-06, 9.705903357826173e-05, 0.019113218411803246
INFO - 27.self_attn.q_proj 1st error time: 2.303541421890259
INFO - 27.self_attn.q_proj proxy error after VPTQ: 1.310478410232463e-06, 9.705903357826173e-05, 0.013501869514584541
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3590.42138671875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 27.self_attn.q_proj residual time: 0.3180427551269531
INFO - 27.self_attn.q_proj 2ed gptq time: 2.599459171295166
INFO - 27.self_attn.q_proj 2ed error time: 0.00039958953857421875
INFO - 27.self_attn.q_proj proxy error after residual VPTQ: 4.6006817910893005e-07, 9.705903357826173e-05, 0.004740085918456316
INFO - load Hessian from /workspace/hessians/quip-format/hessians/27_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:08:04 27.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:27.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 28 iterations, error 2040.8125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 27.self_attn.k_proj 1st kmeans time: 2.8243017196655273
INFO - 27.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 27.self_attn.k_proj proxy error before VPTQ: 4.9370942178939e-07, 2.9934868507552892e-05, 0.01649278774857521
INFO - 27.self_attn.k_proj 1st error time: 1.5492019653320312
INFO - 27.self_attn.k_proj proxy error after VPTQ: 3.9840250565248425e-07, 2.9934868507552892e-05, 0.013308977708220482
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 878.19677734375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 27.self_attn.k_proj residual time: 0.0905451774597168
INFO - 27.self_attn.k_proj 2ed gptq time: 2.0652682781219482
INFO - 27.self_attn.k_proj 2ed error time: 0.0005006790161132812
INFO - 27.self_attn.k_proj proxy error after residual VPTQ: 1.3578912216871686e-07, 2.9934868507552892e-05, 0.004536152351647615
INFO - load Hessian from /workspace/hessians/quip-format/hessians/27_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:08:12 27.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:27.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 23 iterations, error 2842.589599609375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 27.self_attn.v_proj 1st kmeans time: 2.3584840297698975
INFO - 27.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 27.self_attn.v_proj proxy error before VPTQ: 1.643708174015046e-07, 5.5578339015482925e-06, 0.029574617743492126
INFO - 27.self_attn.v_proj 1st error time: 1.5344150066375732
INFO - 27.self_attn.v_proj proxy error after VPTQ: 1.2776172297890298e-07, 5.5578339015482925e-06, 0.02298768237233162
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1166.5743408203125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 27.self_attn.v_proj residual time: 0.09457635879516602
INFO - 27.self_attn.v_proj 2ed gptq time: 2.0685272216796875
INFO - 27.self_attn.v_proj 2ed error time: 0.0005133152008056641
INFO - 27.self_attn.v_proj proxy error after residual VPTQ: 4.1544645057456364e-08, 5.5578339015482925e-06, 0.007474970538169146
INFO - load Hessian from /workspace/hessians/quip-format/hessians/27_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:08:19 27.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:27.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 65 iterations, error 19064.40234375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 27.self_attn.o_proj 1st kmeans time: 25.191309452056885
INFO - 27.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 27.self_attn.o_proj proxy error before VPTQ: 1.6670719560352154e-06, 0.00012937553401570767, 0.01288552675396204
INFO - 27.self_attn.o_proj 1st error time: 2.368607997894287
INFO - 27.self_attn.o_proj proxy error after VPTQ: 9.008637107399409e-07, 0.00012937553401570767, 0.006963169202208519
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7273.1220703125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 27.self_attn.o_proj residual time: 0.3138132095336914
INFO - 27.self_attn.o_proj 2ed gptq time: 2.651472568511963
INFO - 27.self_attn.o_proj 2ed error time: 0.0011742115020751953
INFO - 27.self_attn.o_proj proxy error after residual VPTQ: 2.896853175116121e-07, 0.00012937553401570767, 0.0022391043603420258
INFO - load Hessian from /workspace/hessians/quip-format/hessians/27_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:08:52 27.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:27.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 67334.28125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 27.mlp.gate_proj 1st kmeans time: 132.40029978752136
INFO - 27.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 27.mlp.gate_proj proxy error before VPTQ: 1.0067228686239105e-05, 0.00019832394900731742, 0.05076153948903084
INFO - 27.mlp.gate_proj 1st error time: 4.827641725540161
INFO - 27.mlp.gate_proj proxy error after VPTQ: 7.886678758950438e-06, 0.00019832394900731742, 0.039766646921634674
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 24285.330078125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 27.mlp.gate_proj residual time: 1.0812897682189941
INFO - 27.mlp.gate_proj 2ed gptq time: 5.055383682250977
INFO - 27.mlp.gate_proj 2ed error time: 0.0013897418975830078
INFO - 27.mlp.gate_proj proxy error after residual VPTQ: 2.555218316047103e-06, 0.00019832394900731742, 0.012884063646197319
INFO - load Hessian from /workspace/hessians/quip-format/hessians/27_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:11:17 27.mlp.up_proj
INFO - enabling norm dim 0, layer_name:27.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 68033.609375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 27.mlp.up_proj 1st kmeans time: 132.90456080436707
INFO - 27.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 27.mlp.up_proj proxy error before VPTQ: 9.243409294867888e-06, 0.00016210964531637728, 0.057019490748643875
INFO - 27.mlp.up_proj 1st error time: 4.807687044143677
INFO - 27.mlp.up_proj proxy error after VPTQ: 7.238130365294637e-06, 0.00016210964531637728, 0.04464959725737572
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 24417.3359375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 27.mlp.up_proj residual time: 1.032738208770752
INFO - 27.mlp.up_proj 2ed gptq time: 5.0634987354278564
INFO - 27.mlp.up_proj 2ed error time: 0.0012202262878417969
INFO - 27.mlp.up_proj proxy error after residual VPTQ: 2.339911134185968e-06, 0.00016210964531637728, 0.01443412620574236
INFO - load Hessian from /workspace/hessians/quip-format/hessians/27_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:13:45 27.mlp.down_proj
INFO - enabling norm dim 0, layer_name:27.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 87433.2265625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 27.mlp.down_proj 1st kmeans time: 132.38586854934692
INFO - 27.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 27.mlp.down_proj proxy error before VPTQ: 7.839806812626193e-07, 2.0200575818307698e-05, 0.03880981728434563
INFO - 27.mlp.down_proj 1st error time: 8.958220958709717
INFO - 27.mlp.down_proj proxy error after VPTQ: 6.511475589832116e-07, 2.0200575818307698e-05, 0.032234109938144684
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 30396.181640625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 27.mlp.down_proj residual time: 1.0070598125457764
INFO - 27.mlp.down_proj 2ed gptq time: 9.637609958648682
INFO - 27.mlp.down_proj 2ed error time: 0.002543926239013672
INFO - 27.mlp.down_proj proxy error after residual VPTQ: 2.1042544062765955e-07, 2.0200575818307698e-05, 0.010416803881525993
INFO - ----Quantizing layer 28 ...---- 2026-02-23 12:16:22 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/28_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:16:22 28.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:28.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 75 iterations, error 8738.509765625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 28.self_attn.q_proj 1st kmeans time: 28.650121450424194
INFO - 28.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 28.self_attn.q_proj proxy error before VPTQ: 1.692627847660333e-06, 0.0001164740533567965, 0.014532230794429779
INFO - 28.self_attn.q_proj 1st error time: 2.350775957107544
INFO - 28.self_attn.q_proj proxy error after VPTQ: 1.1398997230571695e-06, 0.0001164740533567965, 0.009786726906895638
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3623.34716796875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 28.self_attn.q_proj residual time: 0.31698107719421387
INFO - 28.self_attn.q_proj 2ed gptq time: 2.607013463973999
INFO - 28.self_attn.q_proj 2ed error time: 0.00035262107849121094
INFO - 28.self_attn.q_proj proxy error after residual VPTQ: 3.927915770418622e-07, 0.0001164740533567965, 0.003372352570295334
INFO - load Hessian from /workspace/hessians/quip-format/hessians/28_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:16:58 28.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:28.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 25 iterations, error 1912.5994873046875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 28.self_attn.k_proj 1st kmeans time: 2.6309149265289307
INFO - 28.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 28.self_attn.k_proj proxy error before VPTQ: 4.777251660925685e-07, 3.831874346360564e-05, 0.012467140331864357
INFO - 28.self_attn.k_proj 1st error time: 1.5484275817871094
INFO - 28.self_attn.k_proj proxy error after VPTQ: 3.774053425331658e-07, 3.831874346360564e-05, 0.009849105030298233
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 819.8104248046875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 28.self_attn.k_proj residual time: 0.10392880439758301
INFO - 28.self_attn.k_proj 2ed gptq time: 2.073234796524048
INFO - 28.self_attn.k_proj 2ed error time: 0.0005478858947753906
INFO - 28.self_attn.k_proj proxy error after residual VPTQ: 1.2679613803356915e-07, 3.831874346360564e-05, 0.0033089846838265657
INFO - load Hessian from /workspace/hessians/quip-format/hessians/28_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:17:05 28.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:28.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 2537.221435546875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 28.self_attn.v_proj 1st kmeans time: 2.2720015048980713
INFO - 28.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 28.self_attn.v_proj proxy error before VPTQ: 1.5802606867509894e-07, 5.9699168559745885e-06, 0.026470396667718887
INFO - 28.self_attn.v_proj 1st error time: 1.574784755706787
INFO - 28.self_attn.v_proj proxy error after VPTQ: 1.2061074983193976e-07, 5.9699168559745885e-06, 0.020203087478876114
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1055.2607421875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 28.self_attn.v_proj residual time: 0.09439921379089355
INFO - 28.self_attn.v_proj 2ed gptq time: 2.070549249649048
INFO - 28.self_attn.v_proj 2ed error time: 0.0004055500030517578
INFO - 28.self_attn.v_proj proxy error after residual VPTQ: 3.931564052095382e-08, 5.9699168559745885e-06, 0.006585625931620598
INFO - load Hessian from /workspace/hessians/quip-format/hessians/28_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:17:13 28.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:28.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 71 iterations, error 18278.146484375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 28.self_attn.o_proj 1st kmeans time: 27.178409337997437
INFO - 28.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 28.self_attn.o_proj proxy error before VPTQ: 1.7368757880831254e-06, 5.573257294599898e-05, 0.031164463609457016
INFO - 28.self_attn.o_proj 1st error time: 2.40474796295166
INFO - 28.self_attn.o_proj proxy error after VPTQ: 9.614198006602237e-07, 5.573257294599898e-05, 0.01725059002637863
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7058.3388671875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 28.self_attn.o_proj residual time: 0.3212721347808838
INFO - 28.self_attn.o_proj 2ed gptq time: 2.6119773387908936
INFO - 28.self_attn.o_proj 2ed error time: 0.0012111663818359375
INFO - 28.self_attn.o_proj proxy error after residual VPTQ: 3.1050535653776024e-07, 5.573257294599898e-05, 0.005571344401687384
INFO - load Hessian from /workspace/hessians/quip-format/hessians/28_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:17:47 28.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:28.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 63235.28125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 28.mlp.gate_proj 1st kmeans time: 132.42154026031494
INFO - 28.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 28.mlp.gate_proj proxy error before VPTQ: 9.890923138300423e-06, 0.0001931044098455459, 0.051220595836639404
INFO - 28.mlp.gate_proj 1st error time: 5.022355079650879
INFO - 28.mlp.gate_proj proxy error after VPTQ: 7.723167072981596e-06, 0.0001931044098455459, 0.039994772523641586
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 23019.921875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 28.mlp.gate_proj residual time: 1.0275866985321045
INFO - 28.mlp.gate_proj 2ed gptq time: 5.135563135147095
INFO - 28.mlp.gate_proj 2ed error time: 0.0013470649719238281
INFO - 28.mlp.gate_proj proxy error after residual VPTQ: 2.525128365959972e-06, 0.0001931044098455459, 0.013076492585241795
INFO - load Hessian from /workspace/hessians/quip-format/hessians/28_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:20:13 28.mlp.up_proj
INFO - enabling norm dim 0, layer_name:28.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 65502.30078125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 28.mlp.up_proj 1st kmeans time: 132.40867114067078
INFO - 28.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 28.mlp.up_proj proxy error before VPTQ: 8.906917173590045e-06, 0.0001569432788528502, 0.056752461940050125
INFO - 28.mlp.up_proj 1st error time: 4.858876943588257
INFO - 28.mlp.up_proj proxy error after VPTQ: 6.912046956131235e-06, 0.0001569432788528502, 0.04404168948531151
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 23712.8125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 28.mlp.up_proj residual time: 1.0325994491577148
INFO - 28.mlp.up_proj 2ed gptq time: 5.109675884246826
INFO - 28.mlp.up_proj 2ed error time: 0.0011320114135742188
INFO - 28.mlp.up_proj proxy error after residual VPTQ: 2.245853465865366e-06, 0.0001569432788528502, 0.014309968799352646
INFO - load Hessian from /workspace/hessians/quip-format/hessians/28_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:22:41 28.mlp.down_proj
INFO - enabling norm dim 0, layer_name:28.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 85057.9140625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 28.mlp.down_proj 1st kmeans time: 132.4328715801239
INFO - 28.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 28.mlp.down_proj proxy error before VPTQ: 7.863349082981586e-07, 2.227391814813018e-05, 0.03530294448137283
INFO - 28.mlp.down_proj 1st error time: 8.963643550872803
INFO - 28.mlp.down_proj proxy error after VPTQ: 6.558061613759492e-07, 2.227391814813018e-05, 0.029442783445119858
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 29686.0625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 28.mlp.down_proj residual time: 1.0264086723327637
INFO - 28.mlp.down_proj 2ed gptq time: 9.674163103103638
INFO - 28.mlp.down_proj 2ed error time: 0.0029458999633789062
INFO - 28.mlp.down_proj proxy error after residual VPTQ: 2.1277733708302549e-07, 2.227391814813018e-05, 0.009552757255733013
INFO - ----Quantizing layer 29 ...---- 2026-02-23 12:25:17 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/29_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:25:18 29.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:29.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 83 iterations, error 8636.62890625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 29.self_attn.q_proj 1st kmeans time: 31.673768043518066
INFO - 29.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 29.self_attn.q_proj proxy error before VPTQ: 1.7245474737137556e-06, 7.958820788189769e-05, 0.02166838012635708
INFO - 29.self_attn.q_proj 1st error time: 2.3083693981170654
INFO - 29.self_attn.q_proj proxy error after VPTQ: 1.2182349564682227e-06, 7.958820788189769e-05, 0.015306727029383183
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3602.327880859375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 29.self_attn.q_proj residual time: 0.3132288455963135
INFO - 29.self_attn.q_proj 2ed gptq time: 2.6747214794158936
INFO - 29.self_attn.q_proj 2ed error time: 0.0003578662872314453
INFO - 29.self_attn.q_proj proxy error after residual VPTQ: 4.2632836994016543e-07, 7.958820788189769e-05, 0.005356677807867527
INFO - load Hessian from /workspace/hessians/quip-format/hessians/29_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:25:57 29.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:29.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 23 iterations, error 1937.1279296875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 29.self_attn.k_proj 1st kmeans time: 2.3556416034698486
INFO - 29.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 29.self_attn.k_proj proxy error before VPTQ: 4.741644659134181e-07, 2.9944074412924238e-05, 0.015835002064704895
INFO - 29.self_attn.k_proj 1st error time: 1.5274875164031982
INFO - 29.self_attn.k_proj proxy error after VPTQ: 3.8427111803684966e-07, 2.9944074412924238e-05, 0.012832960113883018
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 841.3883056640625
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 29.self_attn.k_proj residual time: 0.09116220474243164
INFO - 29.self_attn.k_proj 2ed gptq time: 2.0470211505889893
INFO - 29.self_attn.k_proj 2ed error time: 0.0005846023559570312
INFO - 29.self_attn.k_proj proxy error after residual VPTQ: 1.311559145733554e-07, 2.9944074412924238e-05, 0.004380029160529375
INFO - load Hessian from /workspace/hessians/quip-format/hessians/29_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:26:04 29.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:29.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 24 iterations, error 2796.283935546875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 29.self_attn.v_proj 1st kmeans time: 2.4535090923309326
INFO - 29.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 29.self_attn.v_proj proxy error before VPTQ: 1.976566466055374e-07, 8.325638191308826e-06, 0.023740720003843307
INFO - 29.self_attn.v_proj 1st error time: 1.4993507862091064
INFO - 29.self_attn.v_proj proxy error after VPTQ: 1.4908796686086134e-07, 8.325638191308826e-06, 0.017907092347741127
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 95 iterations, error 1153.6923828125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 29.self_attn.v_proj residual time: 0.08553838729858398
INFO - 29.self_attn.v_proj 2ed gptq time: 1.9779114723205566
INFO - 29.self_attn.v_proj 2ed error time: 0.0004901885986328125
INFO - 29.self_attn.v_proj proxy error after residual VPTQ: 4.854635804463214e-08, 8.325638191308826e-06, 0.005830947309732437
INFO - load Hessian from /workspace/hessians/quip-format/hessians/29_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:26:11 29.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:29.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 69 iterations, error 17364.119140625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 29.self_attn.o_proj 1st kmeans time: 26.41703963279724
INFO - 29.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 29.self_attn.o_proj proxy error before VPTQ: 2.0736977148771985e-06, 8.05701274657622e-05, 0.02573779784142971
INFO - 29.self_attn.o_proj 1st error time: 2.3848400115966797
INFO - 29.self_attn.o_proj proxy error after VPTQ: 1.316096472692152e-06, 8.05701274657622e-05, 0.016334794461727142
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 6706.89208984375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 29.self_attn.o_proj residual time: 0.31121206283569336
INFO - 29.self_attn.o_proj 2ed gptq time: 2.600017547607422
INFO - 29.self_attn.o_proj 2ed error time: 0.0010232925415039062
INFO - 29.self_attn.o_proj proxy error after residual VPTQ: 4.333214747020975e-07, 8.05701274657622e-05, 0.00537819042801857
INFO - load Hessian from /workspace/hessians/quip-format/hessians/29_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:26:45 29.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:29.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 60882.5234375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 29.mlp.gate_proj 1st kmeans time: 132.3997642993927
INFO - 29.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 29.mlp.gate_proj proxy error before VPTQ: 1.055797292792704e-05, 0.0002137855626642704, 0.049385808408260345
INFO - 29.mlp.gate_proj 1st error time: 4.855961561203003
INFO - 29.mlp.gate_proj proxy error after VPTQ: 8.50539163366193e-06, 0.0002137855626642704, 0.039784688502550125
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 22239.953125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 29.mlp.gate_proj residual time: 1.0575451850891113
INFO - 29.mlp.gate_proj 2ed gptq time: 5.060445308685303
INFO - 29.mlp.gate_proj 2ed error time: 0.0013091564178466797
INFO - 29.mlp.gate_proj proxy error after residual VPTQ: 2.804000587275368e-06, 0.0002137855626642704, 0.013115948997437954
INFO - load Hessian from /workspace/hessians/quip-format/hessians/29_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:29:11 29.mlp.up_proj
INFO - enabling norm dim 0, layer_name:29.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 64887.234375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 29.mlp.up_proj 1st kmeans time: 132.70629453659058
INFO - 29.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 29.mlp.up_proj proxy error before VPTQ: 9.035725270223338e-06, 0.00019979733042418957, 0.04522445425391197
INFO - 29.mlp.up_proj 1st error time: 4.842922687530518
INFO - 29.mlp.up_proj proxy error after VPTQ: 7.006507075857371e-06, 0.00019979733042418957, 0.035068072378635406
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 23513.94921875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 29.mlp.up_proj residual time: 1.0382885932922363
INFO - 29.mlp.up_proj 2ed gptq time: 5.060592889785767
INFO - 29.mlp.up_proj 2ed error time: 0.0011525154113769531
INFO - 29.mlp.up_proj proxy error after residual VPTQ: 2.284093170601409e-06, 0.00019979733042418957, 0.011432050727307796
INFO - load Hessian from /workspace/hessians/quip-format/hessians/29_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:31:38 29.mlp.down_proj
INFO - enabling norm dim 0, layer_name:29.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 84650.40625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 29.mlp.down_proj 1st kmeans time: 132.45607161521912
INFO - 29.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 29.mlp.down_proj proxy error before VPTQ: 7.930556193969096e-07, 2.1461795768118463e-05, 0.03695197030901909
INFO - 29.mlp.down_proj 1st error time: 8.874966621398926
INFO - 29.mlp.down_proj proxy error after VPTQ: 6.666996910098533e-07, 2.1461795768118463e-05, 0.031064487993717194
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 29534.322265625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 29.mlp.down_proj residual time: 1.0800540447235107
INFO - 29.mlp.down_proj 2ed gptq time: 9.648859739303589
INFO - 29.mlp.down_proj 2ed error time: 0.0023577213287353516
INFO - 29.mlp.down_proj proxy error after residual VPTQ: 2.1688556728349795e-07, 2.1461795768118463e-05, 0.010105658322572708
INFO - ----Quantizing layer 30 ...---- 2026-02-23 12:34:15 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/30_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:34:16 30.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:30.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 67 iterations, error 8060.47314453125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 30.self_attn.q_proj 1st kmeans time: 25.662317991256714
INFO - 30.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 30.self_attn.q_proj proxy error before VPTQ: 1.7246956076633069e-06, 0.00010070328426081687, 0.017126508057117462
INFO - 30.self_attn.q_proj 1st error time: 2.3098247051239014
INFO - 30.self_attn.q_proj proxy error after VPTQ: 1.2664166888498585e-06, 0.00010070328426081687, 0.01257572416216135
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3390.698486328125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 30.self_attn.q_proj residual time: 0.31681323051452637
INFO - 30.self_attn.q_proj 2ed gptq time: 2.5989880561828613
INFO - 30.self_attn.q_proj 2ed error time: 0.00030732154846191406
INFO - 30.self_attn.q_proj proxy error after residual VPTQ: 4.4803886112276814e-07, 0.00010070328426081687, 0.004449098836630583
INFO - load Hessian from /workspace/hessians/quip-format/hessians/30_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:34:48 30.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:30.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 23 iterations, error 1804.679931640625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 30.self_attn.k_proj 1st kmeans time: 2.3566198348999023
INFO - 30.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 30.self_attn.k_proj proxy error before VPTQ: 4.451713948583347e-07, 2.9315917345229536e-05, 0.015185313299298286
INFO - 30.self_attn.k_proj 1st error time: 1.5431663990020752
INFO - 30.self_attn.k_proj proxy error after VPTQ: 3.6672199144049955e-07, 2.9315917345229536e-05, 0.01250931341201067
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 786.2333374023438
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 30.self_attn.k_proj residual time: 0.09479761123657227
INFO - 30.self_attn.k_proj 2ed gptq time: 2.0511550903320312
INFO - 30.self_attn.k_proj 2ed error time: 0.00038909912109375
INFO - 30.self_attn.k_proj proxy error after residual VPTQ: 1.263620674762933e-07, 2.9315917345229536e-05, 0.004310356918722391
INFO - load Hessian from /workspace/hessians/quip-format/hessians/30_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:34:56 30.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:30.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 23 iterations, error 2610.14453125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 30.self_attn.v_proj 1st kmeans time: 2.3584303855895996
INFO - 30.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 30.self_attn.v_proj proxy error before VPTQ: 1.6154240256582852e-07, 6.093859155953396e-06, 0.02650904841721058
INFO - 30.self_attn.v_proj 1st error time: 1.5449905395507812
INFO - 30.self_attn.v_proj proxy error after VPTQ: 1.24232485632092e-07, 6.093859155953396e-06, 0.020386504009366035
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1084.28662109375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 30.self_attn.v_proj residual time: 0.09373044967651367
INFO - 30.self_attn.v_proj 2ed gptq time: 2.062347412109375
INFO - 30.self_attn.v_proj 2ed error time: 0.00045418739318847656
INFO - 30.self_attn.v_proj proxy error after residual VPTQ: 4.093026007012668e-08, 6.093859155953396e-06, 0.006716640200465918
INFO - load Hessian from /workspace/hessians/quip-format/hessians/30_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:35:04 30.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:30.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 63 iterations, error 17522.962890625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 30.self_attn.o_proj 1st kmeans time: 24.235552549362183
INFO - 30.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 30.self_attn.o_proj proxy error before VPTQ: 1.5813150184840197e-06, 4.815919965039939e-05, 0.032835159450769424
INFO - 30.self_attn.o_proj 1st error time: 2.3139615058898926
INFO - 30.self_attn.o_proj proxy error after VPTQ: 9.362130981571681e-07, 4.815919965039939e-05, 0.01943996362388134
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 6759.87109375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 30.self_attn.o_proj residual time: 0.3143124580383301
INFO - 30.self_attn.o_proj 2ed gptq time: 2.6033496856689453
INFO - 30.self_attn.o_proj 2ed error time: 0.0010323524475097656
INFO - 30.self_attn.o_proj proxy error after residual VPTQ: 3.0440082809946034e-07, 4.815919965039939e-05, 0.006320720072835684
INFO - load Hessian from /workspace/hessians/quip-format/hessians/30_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:35:35 30.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:30.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 62922.27734375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 30.mlp.gate_proj 1st kmeans time: 132.59883284568787
INFO - 30.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 30.mlp.gate_proj proxy error before VPTQ: 1.0098571692651603e-05, 0.00023536212393082678, 0.04290653020143509
INFO - 30.mlp.gate_proj 1st error time: 4.996938228607178
INFO - 30.mlp.gate_proj proxy error after VPTQ: 8.071567208389752e-06, 0.00023536212393082678, 0.0342942476272583
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 22798.373046875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 30.mlp.gate_proj residual time: 1.114565134048462
INFO - 30.mlp.gate_proj 2ed gptq time: 5.03369665145874
INFO - 30.mlp.gate_proj 2ed error time: 0.0014476776123046875
INFO - 30.mlp.gate_proj proxy error after residual VPTQ: 2.6451677967997966e-06, 0.00023536212393082678, 0.011238714680075645
INFO - load Hessian from /workspace/hessians/quip-format/hessians/30_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:38:01 30.mlp.up_proj
INFO - enabling norm dim 0, layer_name:30.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 65856.3203125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 30.mlp.up_proj 1st kmeans time: 132.40529489517212
INFO - 30.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 30.mlp.up_proj proxy error before VPTQ: 9.002904334920458e-06, 0.00014984543668106198, 0.06008126959204674
INFO - 30.mlp.up_proj 1st error time: 4.743911266326904
INFO - 30.mlp.up_proj proxy error after VPTQ: 7.045562597340904e-06, 0.00014984543668106198, 0.0470188669860363
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 98 iterations, error 23760.095703125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 30.mlp.up_proj residual time: 1.038221836090088
INFO - 30.mlp.up_proj 2ed gptq time: 5.065167427062988
INFO - 30.mlp.up_proj 2ed error time: 0.0013494491577148438
INFO - 30.mlp.up_proj proxy error after residual VPTQ: 2.2937674657441676e-06, 0.00014984543668106198, 0.01530755590647459
INFO - load Hessian from /workspace/hessians/quip-format/hessians/30_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:40:28 30.mlp.down_proj
INFO - enabling norm dim 0, layer_name:30.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 84784.2890625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 30.mlp.down_proj 1st kmeans time: 132.6717004776001
INFO - 30.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 30.mlp.down_proj proxy error before VPTQ: 7.457836090907222e-07, 1.6108015188365243e-05, 0.04629891365766525
INFO - 30.mlp.down_proj 1st error time: 8.554127216339111
INFO - 30.mlp.down_proj proxy error after VPTQ: 6.29916087291349e-07, 1.6108015188365243e-05, 0.03910575434565544
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 29551.748046875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 30.mlp.down_proj residual time: 1.021871566772461
INFO - 30.mlp.down_proj 2ed gptq time: 9.47183609008789
INFO - 30.mlp.down_proj 2ed error time: 0.003141641616821289
INFO - 30.mlp.down_proj proxy error after residual VPTQ: 2.0500529274158907e-07, 1.6108015188365243e-05, 0.012726912274956703
INFO - ----Quantizing layer 31 ...---- 2026-02-23 12:43:05 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/31_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:43:06 31.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:31.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 10462.328125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 31.self_attn.q_proj 1st kmeans time: 37.90061378479004
INFO - 31.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 31.self_attn.q_proj proxy error before VPTQ: 1.7208610643137945e-06, 8.357470505870879e-05, 0.02059069275856018
INFO - 31.self_attn.q_proj 1st error time: 2.2819900512695312
INFO - 31.self_attn.q_proj proxy error after VPTQ: 1.2347974234216963e-06, 8.357470505870879e-05, 0.014774774201214314
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 4245.03955078125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 31.self_attn.q_proj residual time: 0.3112325668334961
INFO - 31.self_attn.q_proj 2ed gptq time: 2.621633529663086
INFO - 31.self_attn.q_proj 2ed error time: 0.0003046989440917969
INFO - 31.self_attn.q_proj proxy error after residual VPTQ: 4.217239677473117e-07, 8.357470505870879e-05, 0.0050460719503462315
INFO - load Hessian from /workspace/hessians/quip-format/hessians/31_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:43:50 31.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:31.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 25 iterations, error 2483.769775390625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 31.self_attn.k_proj 1st kmeans time: 2.54464054107666
INFO - 31.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 31.self_attn.k_proj proxy error before VPTQ: 5.020206117478665e-07, 2.87150851363549e-05, 0.0174828190356493
INFO - 31.self_attn.k_proj 1st error time: 1.4921514987945557
INFO - 31.self_attn.k_proj proxy error after VPTQ: 4.0565618064647424e-07, 2.87150851363549e-05, 0.014126935973763466
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1044.8746337890625
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 31.self_attn.k_proj residual time: 0.08936667442321777
INFO - 31.self_attn.k_proj 2ed gptq time: 1.9403369426727295
INFO - 31.self_attn.k_proj 2ed error time: 0.00034618377685546875
INFO - 31.self_attn.k_proj proxy error after residual VPTQ: 1.3461166759043408e-07, 2.87150851363549e-05, 0.004687837790697813
INFO - load Hessian from /workspace/hessians/quip-format/hessians/31_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:43:58 31.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:31.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 21 iterations, error 3003.00244140625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 31.self_attn.v_proj 1st kmeans time: 2.183185338973999
INFO - 31.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 31.self_attn.v_proj proxy error before VPTQ: 1.9367865888852975e-07, 6.823987860116176e-06, 0.028382034972310066
INFO - 31.self_attn.v_proj 1st error time: 1.4947097301483154
INFO - 31.self_attn.v_proj proxy error after VPTQ: 1.5151977095229086e-07, 6.823987860116176e-06, 0.022203991189599037
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1239.0672607421875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 31.self_attn.v_proj residual time: 0.08947348594665527
INFO - 31.self_attn.v_proj 2ed gptq time: 1.9617397785186768
INFO - 31.self_attn.v_proj 2ed error time: 0.0003936290740966797
INFO - 31.self_attn.v_proj proxy error after residual VPTQ: 4.8873872060539725e-08, 6.823987860116176e-06, 0.007162068970501423
INFO - load Hessian from /workspace/hessians/quip-format/hessians/31_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:44:05 31.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:31.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 75 iterations, error 19406.123046875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 31.self_attn.o_proj 1st kmeans time: 28.684069871902466
INFO - 31.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 31.self_attn.o_proj proxy error before VPTQ: 2.279770797031233e-06, 6.319950625766069e-05, 0.03607260435819626
INFO - 31.self_attn.o_proj 1st error time: 2.3024094104766846
INFO - 31.self_attn.o_proj proxy error after VPTQ: 1.387945872011187e-06, 6.319950625766069e-05, 0.021961340680718422
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7364.40283203125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 31.self_attn.o_proj residual time: 0.31493163108825684
INFO - 31.self_attn.o_proj 2ed gptq time: 2.566190481185913
INFO - 31.self_attn.o_proj 2ed error time: 0.0009355545043945312
INFO - 31.self_attn.o_proj proxy error after residual VPTQ: 4.4737976168107707e-07, 6.319950625766069e-05, 0.007078848779201508
INFO - load Hessian from /workspace/hessians/quip-format/hessians/31_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:44:41 31.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:31.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 61734.94921875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 31.mlp.gate_proj 1st kmeans time: 132.4118995666504
INFO - 31.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 31.mlp.gate_proj proxy error before VPTQ: 9.899974429572467e-06, 0.00019118428463116288, 0.05178236588835716
INFO - 31.mlp.gate_proj 1st error time: 4.713281154632568
INFO - 31.mlp.gate_proj proxy error after VPTQ: 7.976053893798962e-06, 0.00019118428463116288, 0.041719190776348114
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 22444.39453125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 31.mlp.gate_proj residual time: 1.0885910987854004
INFO - 31.mlp.gate_proj 2ed gptq time: 5.074794769287109
INFO - 31.mlp.gate_proj 2ed error time: 0.0011739730834960938
INFO - 31.mlp.gate_proj proxy error after residual VPTQ: 2.6247680580127053e-06, 0.00019118428463116288, 0.013728994876146317
INFO - load Hessian from /workspace/hessians/quip-format/hessians/31_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:47:06 31.mlp.up_proj
INFO - enabling norm dim 0, layer_name:31.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 65324.3125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 31.mlp.up_proj 1st kmeans time: 132.4085419178009
INFO - 31.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 31.mlp.up_proj proxy error before VPTQ: 8.8842198238126e-06, 0.00015170489496085793, 0.0585625134408474
INFO - 31.mlp.up_proj 1st error time: 4.709762334823608
INFO - 31.mlp.up_proj proxy error after VPTQ: 6.995214789640158e-06, 0.00015170489496085793, 0.04611067473888397
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 90 iterations, error 23587.892578125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 31.mlp.up_proj residual time: 0.9812827110290527
INFO - 31.mlp.up_proj 2ed gptq time: 5.0150368213653564
INFO - 31.mlp.up_proj 2ed error time: 0.0010325908660888672
INFO - 31.mlp.up_proj proxy error after residual VPTQ: 2.2794010874349624e-06, 0.00015170489496085793, 0.015025231055915356
INFO - load Hessian from /workspace/hessians/quip-format/hessians/31_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:49:34 31.mlp.down_proj
INFO - enabling norm dim 0, layer_name:31.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 83581.7265625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 31.mlp.down_proj 1st kmeans time: 132.38460850715637
INFO - 31.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 31.mlp.down_proj proxy error before VPTQ: 7.371168635472714e-07, 1.757414975145366e-05, 0.04194324463605881
INFO - 31.mlp.down_proj 1st error time: 8.433585166931152
INFO - 31.mlp.down_proj proxy error after VPTQ: 6.272491646086564e-07, 1.757414975145366e-05, 0.035691577941179276
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 29148.45703125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 31.mlp.down_proj residual time: 1.0169389247894287
INFO - 31.mlp.down_proj 2ed gptq time: 9.957844257354736
INFO - 31.mlp.down_proj 2ed error time: 0.0023114681243896484
INFO - 31.mlp.down_proj proxy error after residual VPTQ: 2.0480065643369016e-07, 1.757414975145366e-05, 0.011653517372906208
INFO - ----Quantizing layer 32 ...---- 2026-02-23 12:52:10 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/32_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:52:10 32.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:32.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 81 iterations, error 8892.1904296875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 32.self_attn.q_proj 1st kmeans time: 31.116297006607056
INFO - 32.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 32.self_attn.q_proj proxy error before VPTQ: 1.5838511444599135e-06, 7.060602365527302e-05, 0.02243223786354065
INFO - 32.self_attn.q_proj 1st error time: 2.359182834625244
INFO - 32.self_attn.q_proj proxy error after VPTQ: 1.1466358955658507e-06, 7.060602365527302e-05, 0.016239916905760765
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3715.48876953125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 32.self_attn.q_proj residual time: 0.3114962577819824
INFO - 32.self_attn.q_proj 2ed gptq time: 2.615068197250366
INFO - 32.self_attn.q_proj 2ed error time: 0.0003781318664550781
INFO - 32.self_attn.q_proj proxy error after residual VPTQ: 4.0241567944576673e-07, 7.060602365527302e-05, 0.005699452478438616
INFO - load Hessian from /workspace/hessians/quip-format/hessians/32_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:52:48 32.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:32.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 27 iterations, error 2112.287841796875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 32.self_attn.k_proj 1st kmeans time: 2.742255926132202
INFO - 32.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 32.self_attn.k_proj proxy error before VPTQ: 3.649819291240419e-07, 2.0349274564068764e-05, 0.017935870215296745
INFO - 32.self_attn.k_proj 1st error time: 1.5371880531311035
INFO - 32.self_attn.k_proj proxy error after VPTQ: 2.9989769245730713e-07, 2.0349274564068764e-05, 0.014737512916326523
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 911.639892578125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 32.self_attn.k_proj residual time: 0.09167265892028809
INFO - 32.self_attn.k_proj 2ed gptq time: 2.055602788925171
INFO - 32.self_attn.k_proj 2ed error time: 0.0004048347473144531
INFO - 32.self_attn.k_proj proxy error after residual VPTQ: 1.0251007864781059e-07, 2.0349274564068764e-05, 0.005037529859691858
INFO - load Hessian from /workspace/hessians/quip-format/hessians/32_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:52:56 32.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:32.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 21 iterations, error 2608.104736328125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 32.self_attn.v_proj 1st kmeans time: 2.176485300064087
INFO - 32.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 32.self_attn.v_proj proxy error before VPTQ: 2.298348533713579e-07, 7.34391051082639e-06, 0.03129597753286362
INFO - 32.self_attn.v_proj 1st error time: 1.5434343814849854
INFO - 32.self_attn.v_proj proxy error after VPTQ: 1.842059020873421e-07, 7.34391051082639e-06, 0.02508280798792839
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1105.89453125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 32.self_attn.v_proj residual time: 0.09065413475036621
INFO - 32.self_attn.v_proj 2ed gptq time: 2.0515730381011963
INFO - 32.self_attn.v_proj 2ed error time: 0.0004115104675292969
INFO - 32.self_attn.v_proj proxy error after residual VPTQ: 6.110153094596171e-08, 7.34391051082639e-06, 0.008320027031004429
INFO - load Hessian from /workspace/hessians/quip-format/hessians/32_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:53:03 32.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:32.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 80 iterations, error 17159.12109375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 32.self_attn.o_proj 1st kmeans time: 30.579280138015747
INFO - 32.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 32.self_attn.o_proj proxy error before VPTQ: 2.07397306439816e-06, 8.729504770599306e-05, 0.02375819906592369
INFO - 32.self_attn.o_proj 1st error time: 2.323863983154297
INFO - 32.self_attn.o_proj proxy error after VPTQ: 1.256402129001799e-06, 8.729504770599306e-05, 0.014392593875527382
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 6586.15478515625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 32.self_attn.o_proj residual time: 0.3127753734588623
INFO - 32.self_attn.o_proj 2ed gptq time: 2.597649097442627
INFO - 32.self_attn.o_proj 2ed error time: 0.0010874271392822266
INFO - 32.self_attn.o_proj proxy error after residual VPTQ: 4.138824749588821e-07, 8.729504770599306e-05, 0.004741190932691097
INFO - load Hessian from /workspace/hessians/quip-format/hessians/32_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:53:41 32.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:32.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 55756.21875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 32.mlp.gate_proj 1st kmeans time: 132.40698719024658
INFO - 32.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 32.mlp.gate_proj proxy error before VPTQ: 1.094411072699586e-05, 0.00029195364913903177, 0.03748578205704689
INFO - 32.mlp.gate_proj 1st error time: 4.772298097610474
INFO - 32.mlp.gate_proj proxy error after VPTQ: 8.96884739631787e-06, 0.00029195364913903177, 0.03072010725736618
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 20484.357421875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 32.mlp.gate_proj residual time: 1.042104721069336
INFO - 32.mlp.gate_proj 2ed gptq time: 5.107747793197632
INFO - 32.mlp.gate_proj 2ed error time: 0.0012578964233398438
INFO - 32.mlp.gate_proj proxy error after residual VPTQ: 2.995437625941122e-06, 0.00029195364913903177, 0.010259976610541344
INFO - load Hessian from /workspace/hessians/quip-format/hessians/32_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:56:07 32.mlp.up_proj
INFO - enabling norm dim 0, layer_name:32.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 62196.8125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 32.mlp.up_proj 1st kmeans time: 132.405433177948
INFO - 32.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 32.mlp.up_proj proxy error before VPTQ: 9.251534720533527e-06, 0.0006153141148388386, 0.01503546629101038
INFO - 32.mlp.up_proj 1st error time: 4.752234697341919
INFO - 32.mlp.up_proj proxy error after VPTQ: 7.231627932924312e-06, 0.0006153141148388386, 0.011752741411328316
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 22606.07421875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 32.mlp.up_proj residual time: 1.0460259914398193
INFO - 32.mlp.up_proj 2ed gptq time: 5.06076455116272
INFO - 32.mlp.up_proj 2ed error time: 0.0011703968048095703
INFO - 32.mlp.up_proj proxy error after residual VPTQ: 2.396477611910086e-06, 0.0006153141148388386, 0.0038947223220020533
INFO - load Hessian from /workspace/hessians/quip-format/hessians/32_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 12:58:34 32.mlp.down_proj
INFO - enabling norm dim 0, layer_name:32.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 81098.953125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 32.mlp.down_proj 1st kmeans time: 132.37167048454285
INFO - 32.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 32.mlp.down_proj proxy error before VPTQ: 7.768542218400398e-07, 2.4746143026277423e-05, 0.03139294311404228
INFO - 32.mlp.down_proj 1st error time: 9.103747129440308
INFO - 32.mlp.down_proj proxy error after VPTQ: 6.453860237343179e-07, 2.4746143026277423e-05, 0.026080267503857613
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 28577.07421875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 32.mlp.down_proj residual time: 1.0539913177490234
INFO - 32.mlp.down_proj 2ed gptq time: 9.685003519058228
INFO - 32.mlp.down_proj 2ed error time: 0.0031731128692626953
INFO - 32.mlp.down_proj proxy error after residual VPTQ: 2.1158759011541406e-07, 2.4746143026277423e-05, 0.008550326339900494
INFO - ----Quantizing layer 33 ...---- 2026-02-23 13:01:11 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/33_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:01:11 33.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:33.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 78 iterations, error 9422.552734375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 33.self_attn.q_proj 1st kmeans time: 29.76014256477356
INFO - 33.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 33.self_attn.q_proj proxy error before VPTQ: 1.708262288957485e-06, 8.063422137638554e-05, 0.02118532545864582
INFO - 33.self_attn.q_proj 1st error time: 2.3094217777252197
INFO - 33.self_attn.q_proj proxy error after VPTQ: 1.258012389371288e-06, 8.063422137638554e-05, 0.01560147013515234
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3918.299072265625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 33.self_attn.q_proj residual time: 0.3156147003173828
INFO - 33.self_attn.q_proj 2ed gptq time: 2.607621908187866
INFO - 33.self_attn.q_proj 2ed error time: 0.0009844303131103516
INFO - 33.self_attn.q_proj proxy error after residual VPTQ: 4.4047425262760953e-07, 8.063422137638554e-05, 0.005462621804326773
INFO - load Hessian from /workspace/hessians/quip-format/hessians/33_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:01:48 33.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:33.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 2162.676513671875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 33.self_attn.k_proj 1st kmeans time: 2.270083427429199
INFO - 33.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 33.self_attn.k_proj proxy error before VPTQ: 3.900475746831944e-07, 2.0491526811383665e-05, 0.019034577533602715
INFO - 33.self_attn.k_proj 1st error time: 1.5547258853912354
INFO - 33.self_attn.k_proj proxy error after VPTQ: 3.250528095577465e-07, 2.0491526811383665e-05, 0.01586279086768627
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 938.6104736328125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 33.self_attn.k_proj residual time: 0.09970521926879883
INFO - 33.self_attn.k_proj 2ed gptq time: 2.0649960041046143
INFO - 33.self_attn.k_proj 2ed error time: 0.0005407333374023438
INFO - 33.self_attn.k_proj proxy error after residual VPTQ: 1.1112354059150675e-07, 2.0491526811383665e-05, 0.005422901827841997
INFO - load Hessian from /workspace/hessians/quip-format/hessians/33_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:01:56 33.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:33.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 3133.390380859375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 33.self_attn.v_proj 1st kmeans time: 2.280320882797241
INFO - 33.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 33.self_attn.v_proj proxy error before VPTQ: 2.351205097284037e-07, 7.039110641926527e-06, 0.03340201824903488
INFO - 33.self_attn.v_proj 1st error time: 1.5216879844665527
INFO - 33.self_attn.v_proj proxy error after VPTQ: 1.7218178527400596e-07, 7.039110641926527e-06, 0.024460729211568832
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1215.7857666015625
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 33.self_attn.v_proj residual time: 0.09975934028625488
INFO - 33.self_attn.v_proj 2ed gptq time: 1.9706482887268066
INFO - 33.self_attn.v_proj 2ed error time: 0.000514984130859375
INFO - 33.self_attn.v_proj proxy error after residual VPTQ: 5.6238313561607356e-08, 7.039110641926527e-06, 0.007989405654370785
INFO - load Hessian from /workspace/hessians/quip-format/hessians/33_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:02:03 33.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:33.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 80 iterations, error 19685.86328125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 33.self_attn.o_proj 1st kmeans time: 30.83019995689392
INFO - 33.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 33.self_attn.o_proj proxy error before VPTQ: 2.015751761064166e-06, 0.0001262557925656438, 0.015965618193149567
INFO - 33.self_attn.o_proj 1st error time: 2.327564239501953
INFO - 33.self_attn.o_proj proxy error after VPTQ: 1.1905590326932725e-06, 0.0001262557925656438, 0.00942973792552948
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7452.12109375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 33.self_attn.o_proj residual time: 0.3194897174835205
INFO - 33.self_attn.o_proj 2ed gptq time: 2.6055495738983154
INFO - 33.self_attn.o_proj 2ed error time: 0.0012128353118896484
INFO - 33.self_attn.o_proj proxy error after residual VPTQ: 3.8231991084103356e-07, 0.0001262557925656438, 0.0030281376093626022
INFO - load Hessian from /workspace/hessians/quip-format/hessians/33_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:02:41 33.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:33.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 57337.21875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 33.mlp.gate_proj 1st kmeans time: 132.4002332687378
INFO - 33.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 33.mlp.gate_proj proxy error before VPTQ: 1.1054882634198293e-05, 0.0002671319234650582, 0.04138360917568207
INFO - 33.mlp.gate_proj 1st error time: 4.777752876281738
INFO - 33.mlp.gate_proj proxy error after VPTQ: 9.051775123225525e-06, 0.0002671319234650582, 0.033885035663843155
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 93 iterations, error 21084.517578125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 33.mlp.gate_proj residual time: 1.0189507007598877
INFO - 33.mlp.gate_proj 2ed gptq time: 5.120579481124878
INFO - 33.mlp.gate_proj 2ed error time: 0.0014450550079345703
INFO - 33.mlp.gate_proj proxy error after residual VPTQ: 3.022237251570914e-06, 0.0002671319234650582, 0.011313650757074356
INFO - load Hessian from /workspace/hessians/quip-format/hessians/33_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:05:06 33.mlp.up_proj
INFO - enabling norm dim 0, layer_name:33.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 64986.359375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 33.mlp.up_proj 1st kmeans time: 132.3998839855194
INFO - 33.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 33.mlp.up_proj proxy error before VPTQ: 8.35737409943249e-06, 0.0001379584864480421, 0.060578905045986176
INFO - 33.mlp.up_proj 1st error time: 4.76986026763916
INFO - 33.mlp.up_proj proxy error after VPTQ: 6.377059889928205e-06, 0.0001379584864480421, 0.04622448608279228
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 23503.87890625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 33.mlp.up_proj residual time: 1.0296974182128906
INFO - 33.mlp.up_proj 2ed gptq time: 5.062307834625244
INFO - 33.mlp.up_proj 2ed error time: 0.0012402534484863281
INFO - 33.mlp.up_proj proxy error after residual VPTQ: 2.084255356749054e-06, 0.0001379584864480421, 0.015107844024896622
INFO - load Hessian from /workspace/hessians/quip-format/hessians/33_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:07:34 33.mlp.down_proj
INFO - enabling norm dim 0, layer_name:33.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 83337.9375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 33.mlp.down_proj 1st kmeans time: 132.57162427902222
INFO - 33.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 33.mlp.down_proj proxy error before VPTQ: 7.102899530764262e-07, 1.8206745153293014e-05, 0.03901246190071106
INFO - 33.mlp.down_proj 1st error time: 8.617006301879883
INFO - 33.mlp.down_proj proxy error after VPTQ: 5.767058155470295e-07, 1.8206745153293014e-05, 0.031675394624471664
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 29252.162109375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 33.mlp.down_proj residual time: 1.0463061332702637
INFO - 33.mlp.down_proj 2ed gptq time: 9.57530665397644
INFO - 33.mlp.down_proj 2ed error time: 0.003188610076904297
INFO - 33.mlp.down_proj proxy error after residual VPTQ: 1.8952316338527453e-07, 1.8206745153293014e-05, 0.010409502312541008
INFO - ----Quantizing layer 34 ...---- 2026-02-23 13:10:10 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/34_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:10:11 34.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:34.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 80 iterations, error 10532.068359375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 34.self_attn.q_proj 1st kmeans time: 30.726479291915894
INFO - 34.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 34.self_attn.q_proj proxy error before VPTQ: 1.643296172915143e-06, 7.021666533546522e-05, 0.023403221741318703
INFO - 34.self_attn.q_proj 1st error time: 2.2641186714172363
INFO - 34.self_attn.q_proj proxy error after VPTQ: 1.1616677966230782e-06, 7.021666533546522e-05, 0.016544045880436897
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 4315.58056640625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 34.self_attn.q_proj residual time: 0.31798362731933594
INFO - 34.self_attn.q_proj 2ed gptq time: 2.5597875118255615
INFO - 34.self_attn.q_proj 2ed error time: 0.0002574920654296875
INFO - 34.self_attn.q_proj proxy error after residual VPTQ: 3.997843691649905e-07, 7.021666533546522e-05, 0.0056935823522508144
INFO - load Hessian from /workspace/hessians/quip-format/hessians/34_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:10:49 34.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:34.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 35 iterations, error 2507.0048828125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 34.self_attn.k_proj 1st kmeans time: 3.553248167037964
INFO - 34.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 34.self_attn.k_proj proxy error before VPTQ: 3.919257665074838e-07, 1.9169796360074542e-05, 0.02044496312737465
INFO - 34.self_attn.k_proj 1st error time: 1.5027976036071777
INFO - 34.self_attn.k_proj proxy error after VPTQ: 3.1840949077377445e-07, 1.9169796360074542e-05, 0.016609957441687584
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1065.039306640625
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 34.self_attn.k_proj residual time: 0.0897212028503418
INFO - 34.self_attn.k_proj 2ed gptq time: 1.962000846862793
INFO - 34.self_attn.k_proj 2ed error time: 0.00047206878662109375
INFO - 34.self_attn.k_proj proxy error after residual VPTQ: 1.0584706444660696e-07, 1.9169796360074542e-05, 0.005521554034203291
INFO - load Hessian from /workspace/hessians/quip-format/hessians/34_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:10:57 34.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:34.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 24 iterations, error 3089.76806640625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 34.self_attn.v_proj 1st kmeans time: 2.4532036781311035
INFO - 34.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 34.self_attn.v_proj proxy error before VPTQ: 2.2271636623827362e-07, 6.1608343457919545e-06, 0.036150358617305756
INFO - 34.self_attn.v_proj 1st error time: 1.4940648078918457
INFO - 34.self_attn.v_proj proxy error after VPTQ: 1.7447126765546273e-07, 6.1608343457919545e-06, 0.028319422155618668
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1273.4129638671875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 34.self_attn.v_proj residual time: 0.08831262588500977
INFO - 34.self_attn.v_proj 2ed gptq time: 1.963752031326294
INFO - 34.self_attn.v_proj 2ed error time: 0.00034236907958984375
INFO - 34.self_attn.v_proj proxy error after residual VPTQ: 5.62172743912015e-08, 6.1608343457919545e-06, 0.009124944917857647
INFO - load Hessian from /workspace/hessians/quip-format/hessians/34_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:11:04 34.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:34.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 66 iterations, error 19948.0390625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 34.self_attn.o_proj 1st kmeans time: 25.491186141967773
INFO - 34.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 34.self_attn.o_proj proxy error before VPTQ: 2.5904475933202775e-06, 0.00021564921189565212, 0.012012321501970291
INFO - 34.self_attn.o_proj 1st error time: 2.3439595699310303
INFO - 34.self_attn.o_proj proxy error after VPTQ: 1.6073246342784842e-06, 0.00021564921189565212, 0.00745342206209898
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7517.6796875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 34.self_attn.o_proj residual time: 0.30893635749816895
INFO - 34.self_attn.o_proj 2ed gptq time: 2.6200811862945557
INFO - 34.self_attn.o_proj 2ed error time: 0.000997304916381836
INFO - 34.self_attn.o_proj proxy error after residual VPTQ: 5.256860049485113e-07, 0.00021564921189565212, 0.0024376902729272842
INFO - load Hessian from /workspace/hessians/quip-format/hessians/34_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:11:37 34.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:34.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 53096.8515625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 34.mlp.gate_proj 1st kmeans time: 132.49498295783997
INFO - 34.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 34.mlp.gate_proj proxy error before VPTQ: 1.0801052667375188e-05, 0.00023215642431750894, 0.04652489349246025
INFO - 34.mlp.gate_proj 1st error time: 4.733076810836792
INFO - 34.mlp.gate_proj proxy error after VPTQ: 8.667172551213298e-06, 0.00023215642431750894, 0.03733333200216293
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 19714.41015625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 34.mlp.gate_proj residual time: 1.0732600688934326
INFO - 34.mlp.gate_proj 2ed gptq time: 5.012913942337036
INFO - 34.mlp.gate_proj 2ed error time: 0.0011267662048339844
INFO - 34.mlp.gate_proj proxy error after residual VPTQ: 2.9107591217325535e-06, 0.00023215642431750894, 0.012537921778857708
INFO - load Hessian from /workspace/hessians/quip-format/hessians/34_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:14:02 34.mlp.up_proj
INFO - enabling norm dim 0, layer_name:34.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 63040.046875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 34.mlp.up_proj 1st kmeans time: 132.40805196762085
INFO - 34.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 34.mlp.up_proj proxy error before VPTQ: 8.293820428662002e-06, 0.0002520769485272467, 0.03290193900465965
INFO - 34.mlp.up_proj 1st error time: 4.762873649597168
INFO - 34.mlp.up_proj proxy error after VPTQ: 6.1662294683628716e-06, 0.0002520769485272467, 0.024461695924401283
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 93 iterations, error 22933.11328125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 34.mlp.up_proj residual time: 1.0289249420166016
INFO - 34.mlp.up_proj 2ed gptq time: 5.02303147315979
INFO - 34.mlp.up_proj 2ed error time: 0.0011892318725585938
INFO - 34.mlp.up_proj proxy error after residual VPTQ: 2.0158395273028873e-06, 0.0002520769485272467, 0.007996921427547932
INFO - load Hessian from /workspace/hessians/quip-format/hessians/34_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:16:30 34.mlp.down_proj
INFO - enabling norm dim 0, layer_name:34.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 81184.6015625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 34.mlp.down_proj 1st kmeans time: 132.56607580184937
INFO - 34.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 34.mlp.down_proj proxy error before VPTQ: 6.752755439265457e-07, 1.652918399486225e-05, 0.040853530168533325
INFO - 34.mlp.down_proj 1st error time: 8.499126672744751
INFO - 34.mlp.down_proj proxy error after VPTQ: 5.272382281873433e-07, 1.652918399486225e-05, 0.0318974144756794
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 28766.71875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 34.mlp.down_proj residual time: 1.045133113861084
INFO - 34.mlp.down_proj 2ed gptq time: 9.573116064071655
INFO - 34.mlp.down_proj 2ed error time: 0.002396106719970703
INFO - 34.mlp.down_proj proxy error after residual VPTQ: 1.7274609831474663e-07, 1.652918399486225e-05, 0.010450975969433784
INFO - ----Quantizing layer 35 ...---- 2026-02-23 13:19:06 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/35_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:19:06 35.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:35.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 65 iterations, error 9323.166015625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 35.self_attn.q_proj 1st kmeans time: 25.111787796020508
INFO - 35.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 35.self_attn.q_proj proxy error before VPTQ: 1.4789135320825153e-06, 8.301634807139635e-05, 0.017814725637435913
INFO - 35.self_attn.q_proj 1st error time: 2.2607271671295166
INFO - 35.self_attn.q_proj proxy error after VPTQ: 9.740003861224977e-07, 8.301634807139635e-05, 0.011732633225619793
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3841.4765625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 35.self_attn.q_proj residual time: 0.31491851806640625
INFO - 35.self_attn.q_proj 2ed gptq time: 2.551076650619507
INFO - 35.self_attn.q_proj 2ed error time: 0.00025916099548339844
INFO - 35.self_attn.q_proj proxy error after residual VPTQ: 3.3642135122136096e-07, 8.301634807139635e-05, 0.004052470903843641
INFO - load Hessian from /workspace/hessians/quip-format/hessians/35_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:19:38 35.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:35.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 27 iterations, error 2078.07958984375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 35.self_attn.k_proj 1st kmeans time: 2.7565910816192627
INFO - 35.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 35.self_attn.k_proj proxy error before VPTQ: 3.2228132340605953e-07, 2.1176419977564365e-05, 0.015218876302242279
INFO - 35.self_attn.k_proj 1st error time: 1.4994146823883057
INFO - 35.self_attn.k_proj proxy error after VPTQ: 2.480727516740444e-07, 2.1176419977564365e-05, 0.011714574880897999
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 896.51806640625
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 35.self_attn.k_proj residual time: 0.1726682186126709
INFO - 35.self_attn.k_proj 2ed gptq time: 1.9438612461090088
INFO - 35.self_attn.k_proj 2ed error time: 0.0003268718719482422
INFO - 35.self_attn.k_proj proxy error after residual VPTQ: 8.320529332195292e-08, 2.1176419977564365e-05, 0.003929148428142071
INFO - load Hessian from /workspace/hessians/quip-format/hessians/35_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:19:46 35.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:35.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 28 iterations, error 2604.78076171875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 35.self_attn.v_proj 1st kmeans time: 2.84470796585083
INFO - 35.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 35.self_attn.v_proj proxy error before VPTQ: 2.1151734586055682e-07, 7.958346941450145e-06, 0.02657805010676384
INFO - 35.self_attn.v_proj 1st error time: 1.5308630466461182
INFO - 35.self_attn.v_proj proxy error after VPTQ: 1.5497538186082238e-07, 7.958346941450145e-06, 0.01947331242263317
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1103.3555908203125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 35.self_attn.v_proj residual time: 0.09482884407043457
INFO - 35.self_attn.v_proj 2ed gptq time: 2.0499258041381836
INFO - 35.self_attn.v_proj 2ed error time: 0.0004589557647705078
INFO - 35.self_attn.v_proj proxy error after residual VPTQ: 5.0817902774724644e-08, 7.958346941450145e-06, 0.0063854847103357315
INFO - load Hessian from /workspace/hessians/quip-format/hessians/35_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:19:54 35.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:35.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 72 iterations, error 19069.630859375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 35.self_attn.o_proj 1st kmeans time: 28.105342388153076
INFO - 35.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 35.self_attn.o_proj proxy error before VPTQ: 2.2037520466255955e-06, 6.010358629282564e-05, 0.0366658978164196
INFO - 35.self_attn.o_proj 1st error time: 2.2796552181243896
INFO - 35.self_attn.o_proj proxy error after VPTQ: 1.3895839856559178e-06, 6.010358629282564e-05, 0.023119818419218063
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7361.63134765625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 35.self_attn.o_proj residual time: 0.30883193016052246
INFO - 35.self_attn.o_proj 2ed gptq time: 2.5615766048431396
INFO - 35.self_attn.o_proj 2ed error time: 0.0009493827819824219
INFO - 35.self_attn.o_proj proxy error after residual VPTQ: 4.490491392061813e-07, 6.010358629282564e-05, 0.007471253629773855
INFO - load Hessian from /workspace/hessians/quip-format/hessians/35_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:20:29 35.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:35.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 50306.1328125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 35.mlp.gate_proj 1st kmeans time: 133.52067112922668
INFO - 35.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 35.mlp.gate_proj proxy error before VPTQ: 1.0219241630693432e-05, 0.000319858081638813, 0.031949300318956375
INFO - 35.mlp.gate_proj 1st error time: 4.8892738819122314
INFO - 35.mlp.gate_proj proxy error after VPTQ: 7.920702046249062e-06, 0.000319858081638813, 0.02476317621767521
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 84 iterations, error 18728.572265625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 35.mlp.gate_proj residual time: 0.9122312068939209
INFO - 35.mlp.gate_proj 2ed gptq time: 5.0866453647613525
INFO - 35.mlp.gate_proj 2ed error time: 0.012931346893310547
INFO - 35.mlp.gate_proj proxy error after residual VPTQ: 2.653625642778934e-06, 0.000319858081638813, 0.008296259678900242
INFO - load Hessian from /workspace/hessians/quip-format/hessians/35_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:22:56 35.mlp.up_proj
INFO - enabling norm dim 0, layer_name:35.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 62946.30859375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 35.mlp.up_proj 1st kmeans time: 133.49523162841797
INFO - 35.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 35.mlp.up_proj proxy error before VPTQ: 8.409101610595826e-06, 0.0001454705052310601, 0.05780623108148575
INFO - 35.mlp.up_proj 1st error time: 4.714248418807983
INFO - 35.mlp.up_proj proxy error after VPTQ: 6.239532922336366e-06, 0.0001454705052310601, 0.042892083525657654
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 22948.87109375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 35.mlp.up_proj residual time: 1.080740213394165
INFO - 35.mlp.up_proj 2ed gptq time: 5.011059045791626
INFO - 35.mlp.up_proj 2ed error time: 0.0010199546813964844
INFO - 35.mlp.up_proj proxy error after residual VPTQ: 2.038318143604556e-06, 0.0001454705052310601, 0.014011899940669537
INFO - load Hessian from /workspace/hessians/quip-format/hessians/35_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:25:24 35.mlp.down_proj
INFO - enabling norm dim 0, layer_name:35.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 80654.171875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 35.mlp.down_proj 1st kmeans time: 132.47646355628967
INFO - 35.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 35.mlp.down_proj proxy error before VPTQ: 7.026085313555086e-07, 1.736211015668232e-05, 0.040467921644449234
INFO - 35.mlp.down_proj 1st error time: 8.408243894577026
INFO - 35.mlp.down_proj proxy error after VPTQ: 5.547209411815857e-07, 1.736211015668232e-05, 0.03195008635520935
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 28576.9140625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 35.mlp.down_proj residual time: 1.0515995025634766
INFO - 35.mlp.down_proj 2ed gptq time: 9.472729206085205
INFO - 35.mlp.down_proj 2ed error time: 0.0023529529571533203
INFO - 35.mlp.down_proj proxy error after residual VPTQ: 1.8253219025154976e-07, 1.736211015668232e-05, 0.010513248853385448
INFO - ----Quantizing layer 36 ...---- 2026-02-23 13:28:00 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/36_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:28:01 36.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:36.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 74 iterations, error 8356.7724609375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 36.self_attn.q_proj 1st kmeans time: 28.38292646408081
INFO - 36.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 36.self_attn.q_proj proxy error before VPTQ: 1.533116119389888e-06, 8.051483746385202e-05, 0.019041411578655243
INFO - 36.self_attn.q_proj 1st error time: 2.3440520763397217
INFO - 36.self_attn.q_proj proxy error after VPTQ: 1.0319525927116047e-06, 8.051483746385202e-05, 0.01281692460179329
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 3536.60400390625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 36.self_attn.q_proj residual time: 0.3141770362854004
INFO - 36.self_attn.q_proj 2ed gptq time: 2.5620574951171875
INFO - 36.self_attn.q_proj 2ed error time: 0.0003345012664794922
INFO - 36.self_attn.q_proj proxy error after residual VPTQ: 3.633037408690143e-07, 8.051483746385202e-05, 0.004512258339673281
INFO - load Hessian from /workspace/hessians/quip-format/hessians/36_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:28:36 36.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:36.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 28 iterations, error 1962.92919921875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 36.self_attn.k_proj 1st kmeans time: 2.835563898086548
INFO - 36.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 36.self_attn.k_proj proxy error before VPTQ: 2.9198713491496164e-07, 1.6994443285511807e-05, 0.017181329429149628
INFO - 36.self_attn.k_proj 1st error time: 1.5098764896392822
INFO - 36.self_attn.k_proj proxy error after VPTQ: 2.316468794560933e-07, 1.6994443285511807e-05, 0.013630742207169533
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 852.658935546875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 36.self_attn.k_proj residual time: 0.09417271614074707
INFO - 36.self_attn.k_proj 2ed gptq time: 1.9652049541473389
INFO - 36.self_attn.k_proj 2ed error time: 0.000492095947265625
INFO - 36.self_attn.k_proj proxy error after residual VPTQ: 7.841117621865124e-08, 1.6994443285511807e-05, 0.004613930359482765
INFO - load Hessian from /workspace/hessians/quip-format/hessians/36_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:28:44 36.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:36.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 23 iterations, error 2505.4267578125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 36.self_attn.v_proj 1st kmeans time: 2.3777692317962646
INFO - 36.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 36.self_attn.v_proj proxy error before VPTQ: 1.969029312931525e-07, 6.454897629737388e-06, 0.03050442412495613
INFO - 36.self_attn.v_proj 1st error time: 1.5174615383148193
INFO - 36.self_attn.v_proj proxy error after VPTQ: 1.512148912752309e-07, 6.454897629737388e-06, 0.02342638187110424
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1055.6881103515625
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 36.self_attn.v_proj residual time: 0.10106134414672852
INFO - 36.self_attn.v_proj 2ed gptq time: 1.9727363586425781
INFO - 36.self_attn.v_proj 2ed error time: 0.00044226646423339844
INFO - 36.self_attn.v_proj proxy error after residual VPTQ: 4.954735288720258e-08, 6.454897629737388e-06, 0.007675931788980961
INFO - load Hessian from /workspace/hessians/quip-format/hessians/36_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:28:51 36.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:36.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 62 iterations, error 18189.638671875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 36.self_attn.o_proj 1st kmeans time: 23.82906699180603
INFO - 36.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 36.self_attn.o_proj proxy error before VPTQ: 2.3558891371067148e-06, 0.00014096626546233892, 0.016712432727217674
INFO - 36.self_attn.o_proj 1st error time: 2.309642791748047
INFO - 36.self_attn.o_proj proxy error after VPTQ: 1.354130290565081e-06, 0.00014096626546233892, 0.00960605964064598
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7133.24853515625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 36.self_attn.o_proj residual time: 0.30965137481689453
INFO - 36.self_attn.o_proj 2ed gptq time: 2.556124448776245
INFO - 36.self_attn.o_proj 2ed error time: 0.0010538101196289062
INFO - 36.self_attn.o_proj proxy error after residual VPTQ: 4.3528086735022953e-07, 0.00014096626546233892, 0.00308783701620996
INFO - load Hessian from /workspace/hessians/quip-format/hessians/36_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:29:22 36.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:36.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 49392.9921875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 36.mlp.gate_proj 1st kmeans time: 132.38959050178528
INFO - 36.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 36.mlp.gate_proj proxy error before VPTQ: 9.22157778404653e-06, 0.00021813205967191607, 0.04227520525455475
INFO - 36.mlp.gate_proj 1st error time: 4.770623445510864
INFO - 36.mlp.gate_proj proxy error after VPTQ: 6.7887681325373705e-06, 0.00021813205967191607, 0.03112228587269783
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 79 iterations, error 18436.890625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 36.mlp.gate_proj residual time: 0.8931944370269775
INFO - 36.mlp.gate_proj 2ed gptq time: 5.083802938461304
INFO - 36.mlp.gate_proj 2ed error time: 0.001210927963256836
INFO - 36.mlp.gate_proj proxy error after residual VPTQ: 2.263387386847171e-06, 0.00021813205967191607, 0.010376225225627422
INFO - load Hessian from /workspace/hessians/quip-format/hessians/36_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:31:47 36.mlp.up_proj
INFO - enabling norm dim 0, layer_name:36.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 62411.00390625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 36.mlp.up_proj 1st kmeans time: 132.40757989883423
INFO - 36.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 36.mlp.up_proj proxy error before VPTQ: 8.100869308691472e-06, 0.00019269548647571355, 0.042039744555950165
INFO - 36.mlp.up_proj 1st error time: 4.788387775421143
INFO - 36.mlp.up_proj proxy error after VPTQ: 5.752662218583282e-06, 0.00019269548647571355, 0.029853641986846924
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 22826.65625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 36.mlp.up_proj residual time: 1.1077697277069092
INFO - 36.mlp.up_proj 2ed gptq time: 5.064833402633667
INFO - 36.mlp.up_proj 2ed error time: 0.0014209747314453125
INFO - 36.mlp.up_proj proxy error after residual VPTQ: 1.8820309151124093e-06, 0.00019269548647571355, 0.009766865521669388
INFO - load Hessian from /workspace/hessians/quip-format/hessians/36_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:34:15 36.mlp.down_proj
INFO - enabling norm dim 0, layer_name:36.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 77756.890625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 36.mlp.down_proj 1st kmeans time: 132.37331581115723
INFO - 36.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 36.mlp.down_proj proxy error before VPTQ: 6.759493089703028e-07, 2.030044015555177e-05, 0.03329727426171303
INFO - 36.mlp.down_proj 1st error time: 8.41007399559021
INFO - 36.mlp.down_proj proxy error after VPTQ: 5.113863039696298e-07, 2.030044015555177e-05, 0.025190897285938263
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 27864.517578125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 36.mlp.down_proj residual time: 1.0536487102508545
INFO - 36.mlp.down_proj 2ed gptq time: 9.447701930999756
INFO - 36.mlp.down_proj 2ed error time: 0.0022802352905273438
INFO - 36.mlp.down_proj proxy error after residual VPTQ: 1.690664106490658e-07, 2.030044015555177e-05, 0.008328214287757874
INFO - ----Quantizing layer 37 ...---- 2026-02-23 13:36:51 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/37_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:36:52 37.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:37.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 68 iterations, error 10709.4833984375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 37.self_attn.q_proj 1st kmeans time: 26.07335662841797
INFO - 37.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 37.self_attn.q_proj proxy error before VPTQ: 1.4938889307813952e-06, 7.301420555450022e-05, 0.020460249856114388
INFO - 37.self_attn.q_proj 1st error time: 2.260319709777832
INFO - 37.self_attn.q_proj proxy error after VPTQ: 9.586096894054208e-07, 7.301420555450022e-05, 0.01312908437103033
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 100 iterations, error 4355.36767578125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 37.self_attn.q_proj residual time: 0.31467604637145996
INFO - 37.self_attn.q_proj 2ed gptq time: 2.553950786590576
INFO - 37.self_attn.q_proj 2ed error time: 0.0002951622009277344
INFO - 37.self_attn.q_proj proxy error after residual VPTQ: 3.278793485606002e-07, 7.301420555450022e-05, 0.0044906241819262505
INFO - load Hessian from /workspace/hessians/quip-format/hessians/37_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:37:25 37.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:37.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 26 iterations, error 2289.357177734375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 37.self_attn.k_proj 1st kmeans time: 2.6795876026153564
INFO - 37.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 37.self_attn.k_proj proxy error before VPTQ: 3.1046823778524413e-07, 1.5583980712108314e-05, 0.019922267645597458
INFO - 37.self_attn.k_proj 1st error time: 1.4945156574249268
INFO - 37.self_attn.k_proj proxy error after VPTQ: 2.3799796622370195e-07, 1.5583980712108314e-05, 0.015271962620317936
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 978.669189453125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 37.self_attn.k_proj residual time: 0.09371185302734375
INFO - 37.self_attn.k_proj 2ed gptq time: 1.9448184967041016
INFO - 37.self_attn.k_proj 2ed error time: 0.0003273487091064453
INFO - 37.self_attn.k_proj proxy error after residual VPTQ: 7.994151474122191e-08, 1.5583980712108314e-05, 0.005129723809659481
INFO - load Hessian from /workspace/hessians/quip-format/hessians/37_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:37:32 37.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:37.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 30 iterations, error 2761.1484375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 37.self_attn.v_proj 1st kmeans time: 3.0374746322631836
INFO - 37.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 37.self_attn.v_proj proxy error before VPTQ: 2.3627849543572665e-07, 7.140850811993005e-06, 0.033088281750679016
INFO - 37.self_attn.v_proj 1st error time: 1.5042550563812256
INFO - 37.self_attn.v_proj proxy error after VPTQ: 1.7891578352191573e-07, 7.140850811993005e-06, 0.02505524829030037
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1149.877685546875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 37.self_attn.v_proj residual time: 0.09657621383666992
INFO - 37.self_attn.v_proj 2ed gptq time: 1.9481072425842285
INFO - 37.self_attn.v_proj 2ed error time: 0.0003609657287597656
INFO - 37.self_attn.v_proj proxy error after residual VPTQ: 5.7670746400617645e-08, 7.140850811993005e-06, 0.00807617325335741
INFO - load Hessian from /workspace/hessians/quip-format/hessians/37_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:37:40 37.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:37.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 95 iterations, error 19892.7734375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 37.self_attn.o_proj 1st kmeans time: 36.14799928665161
INFO - 37.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 37.self_attn.o_proj proxy error before VPTQ: 2.8640647542488296e-06, 0.00017161935102194548, 0.016688471660017967
INFO - 37.self_attn.o_proj 1st error time: 2.267590045928955
INFO - 37.self_attn.o_proj proxy error after VPTQ: 2.019521161855664e-06, 0.00017161935102194548, 0.01176744420081377
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7721.92333984375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 37.self_attn.o_proj residual time: 0.3127021789550781
INFO - 37.self_attn.o_proj 2ed gptq time: 2.5521299839019775
INFO - 37.self_attn.o_proj 2ed error time: 0.0009133815765380859
INFO - 37.self_attn.o_proj proxy error after residual VPTQ: 6.513831181109708e-07, 0.00017161935102194548, 0.003795510856434703
INFO - load Hessian from /workspace/hessians/quip-format/hessians/37_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:38:23 37.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:37.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 51758.88671875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 37.mlp.gate_proj 1st kmeans time: 132.406076669693
INFO - 37.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 37.mlp.gate_proj proxy error before VPTQ: 9.050400876731146e-06, 0.00019291837816126645, 0.046913109719753265
INFO - 37.mlp.gate_proj 1st error time: 4.772350549697876
INFO - 37.mlp.gate_proj proxy error after VPTQ: 6.407040928024799e-06, 0.00019291837816126645, 0.03321114927530289
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 96 iterations, error 19252.66796875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 37.mlp.gate_proj residual time: 1.0384211540222168
INFO - 37.mlp.gate_proj 2ed gptq time: 5.040986776351929
INFO - 37.mlp.gate_proj 2ed error time: 0.0012295246124267578
INFO - 37.mlp.gate_proj proxy error after residual VPTQ: 2.118997144862078e-06, 0.00019291837816126645, 0.01098390482366085
INFO - load Hessian from /workspace/hessians/quip-format/hessians/37_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:40:49 37.mlp.up_proj
INFO - enabling norm dim 0, layer_name:37.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 63177.01171875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 37.mlp.up_proj 1st kmeans time: 132.50134944915771
INFO - 37.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 37.mlp.up_proj proxy error before VPTQ: 7.960839866427705e-06, 0.00015586972585879266, 0.051073674112558365
INFO - 37.mlp.up_proj 1st error time: 4.759248495101929
INFO - 37.mlp.up_proj proxy error after VPTQ: 5.4554498092329595e-06, 0.00015586972585879266, 0.03500005975365639
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 23141.12890625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 37.mlp.up_proj residual time: 1.0354816913604736
INFO - 37.mlp.up_proj 2ed gptq time: 5.036902666091919
INFO - 37.mlp.up_proj 2ed error time: 0.0012102127075195312
INFO - 37.mlp.up_proj proxy error after residual VPTQ: 1.7733700587996282e-06, 0.00015586972585879266, 0.011377258226275444
INFO - load Hessian from /workspace/hessians/quip-format/hessians/37_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:43:17 37.mlp.down_proj
INFO - enabling norm dim 0, layer_name:37.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 78762.765625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 37.mlp.down_proj 1st kmeans time: 132.4732370376587
INFO - 37.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 37.mlp.down_proj proxy error before VPTQ: 6.753079446752963e-07, 2.148189378203824e-05, 0.031436145305633545
INFO - 37.mlp.down_proj 1st error time: 8.677056789398193
INFO - 37.mlp.down_proj proxy error after VPTQ: 4.990696993445454e-07, 2.148189378203824e-05, 0.0232321098446846
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 28204.953125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 37.mlp.down_proj residual time: 1.0837037563323975
INFO - 37.mlp.down_proj 2ed gptq time: 9.462979078292847
INFO - 37.mlp.down_proj 2ed error time: 0.0022590160369873047
INFO - 37.mlp.down_proj proxy error after residual VPTQ: 1.6407412317676062e-07, 2.148189378203824e-05, 0.00763778667896986
INFO - ----Quantizing layer 38 ...---- 2026-02-23 13:45:53 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/38_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:45:54 38.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:38.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 69 iterations, error 10033.5693359375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 38.self_attn.q_proj 1st kmeans time: 26.48872137069702
INFO - 38.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 38.self_attn.q_proj proxy error before VPTQ: 1.4705412922921823e-06, 8.846896525938064e-05, 0.016622114926576614
INFO - 38.self_attn.q_proj 1st error time: 2.335958957672119
INFO - 38.self_attn.q_proj proxy error after VPTQ: 9.588643479219172e-07, 8.846896525938064e-05, 0.01083842571824789
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 4153.74267578125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 38.self_attn.q_proj residual time: 0.35480189323425293
INFO - 38.self_attn.q_proj 2ed gptq time: 2.5936057567596436
INFO - 38.self_attn.q_proj 2ed error time: 0.00047779083251953125
INFO - 38.self_attn.q_proj proxy error after residual VPTQ: 3.3216548445125227e-07, 8.846896525938064e-05, 0.003754599019885063
INFO - load Hessian from /workspace/hessians/quip-format/hessians/38_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:46:27 38.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:38.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 2436.20849609375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 38.self_attn.k_proj 1st kmeans time: 2.3014698028564453
INFO - 38.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 38.self_attn.k_proj proxy error before VPTQ: 2.9154927005947684e-07, 1.570897438796237e-05, 0.01855940744280815
INFO - 38.self_attn.k_proj 1st error time: 1.516721248626709
INFO - 38.self_attn.k_proj proxy error after VPTQ: 2.1893131929573428e-07, 1.570897438796237e-05, 0.013936704024672508
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1040.6951904296875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 38.self_attn.k_proj residual time: 0.09517741203308105
INFO - 38.self_attn.k_proj 2ed gptq time: 1.9729812145233154
INFO - 38.self_attn.k_proj 2ed error time: 0.0003437995910644531
INFO - 38.self_attn.k_proj proxy error after residual VPTQ: 7.301781579371891e-08, 1.570897438796237e-05, 0.004648159258067608
INFO - load Hessian from /workspace/hessians/quip-format/hessians/38_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:46:35 38.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:38.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 3400.20751953125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 38.self_attn.v_proj 1st kmeans time: 2.2675135135650635
INFO - 38.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 38.self_attn.v_proj proxy error before VPTQ: 3.192878637037211e-07, 1.3408483027887996e-05, 0.02381237782537937
INFO - 38.self_attn.v_proj 1st error time: 1.4951579570770264
INFO - 38.self_attn.v_proj proxy error after VPTQ: 1.9975568932295573e-07, 1.3408483027887996e-05, 0.014897709712386131
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 100 iterations, error 1335.97509765625
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 38.self_attn.v_proj residual time: 0.08869600296020508
INFO - 38.self_attn.v_proj 2ed gptq time: 1.9593610763549805
INFO - 38.self_attn.v_proj 2ed error time: 0.000335693359375
INFO - 38.self_attn.v_proj proxy error after residual VPTQ: 6.432304644476972e-08, 1.3408483027887996e-05, 0.004797190427780151
INFO - load Hessian from /workspace/hessians/quip-format/hessians/38_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:46:42 38.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:38.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 63 iterations, error 20106.72265625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 38.self_attn.o_proj 1st kmeans time: 24.20681357383728
INFO - 38.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 38.self_attn.o_proj proxy error before VPTQ: 2.388721213719691e-06, 9.217944170814008e-05, 0.025913817808032036
INFO - 38.self_attn.o_proj 1st error time: 2.3197741508483887
INFO - 38.self_attn.o_proj proxy error after VPTQ: 1.6721277233955334e-06, 9.217944170814008e-05, 0.018139921128749847
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7629.244140625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 38.self_attn.o_proj residual time: 0.3091268539428711
INFO - 38.self_attn.o_proj 2ed gptq time: 2.642639398574829
INFO - 38.self_attn.o_proj 2ed error time: 0.0010356903076171875
INFO - 38.self_attn.o_proj proxy error after residual VPTQ: 5.390658088799682e-07, 9.217944170814008e-05, 0.00584800448268652
INFO - load Hessian from /workspace/hessians/quip-format/hessians/38_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:47:13 38.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:38.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 57327.890625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 38.mlp.gate_proj 1st kmeans time: 132.3977026939392
INFO - 38.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 38.mlp.gate_proj proxy error before VPTQ: 9.056899216375314e-06, 0.00018023682059720159, 0.05024999380111694
INFO - 38.mlp.gate_proj 1st error time: 4.815022230148315
INFO - 38.mlp.gate_proj proxy error after VPTQ: 6.498557922895998e-06, 0.00018023682059720159, 0.036055661737918854
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 21028.962890625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 38.mlp.gate_proj residual time: 1.0914511680603027
INFO - 38.mlp.gate_proj 2ed gptq time: 5.039376735687256
INFO - 38.mlp.gate_proj 2ed error time: 0.0016257762908935547
INFO - 38.mlp.gate_proj proxy error after residual VPTQ: 2.1276366624078946e-06, 0.00018023682059720159, 0.01180467288941145
INFO - load Hessian from /workspace/hessians/quip-format/hessians/38_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:49:39 38.mlp.up_proj
INFO - enabling norm dim 0, layer_name:38.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 67316.0
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 38.mlp.up_proj 1st kmeans time: 132.6884298324585
INFO - 38.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 38.mlp.up_proj proxy error before VPTQ: 7.943053788039833e-06, 0.0001702791778370738, 0.04664723947644234
INFO - 38.mlp.up_proj 1st error time: 4.781322479248047
INFO - 38.mlp.up_proj proxy error after VPTQ: 5.5000973588903435e-06, 0.0001702791778370738, 0.032300468534231186
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 24509.208984375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 38.mlp.up_proj residual time: 1.0993702411651611
INFO - 38.mlp.up_proj 2ed gptq time: 5.023278713226318
INFO - 38.mlp.up_proj 2ed error time: 0.0012319087982177734
INFO - 38.mlp.up_proj proxy error after residual VPTQ: 1.7750871847965755e-06, 0.0001702791778370738, 0.010424570180475712
INFO - load Hessian from /workspace/hessians/quip-format/hessians/38_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:52:07 38.mlp.down_proj
INFO - enabling norm dim 0, layer_name:38.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 81620.8046875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 38.mlp.down_proj 1st kmeans time: 132.45241951942444
INFO - 38.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 38.mlp.down_proj proxy error before VPTQ: 6.440897095671971e-07, 1.9424798665568233e-05, 0.03315811604261398
INFO - 38.mlp.down_proj 1st error time: 8.92317533493042
INFO - 38.mlp.down_proj proxy error after VPTQ: 4.803270599040843e-07, 1.9424798665568233e-05, 0.02472751773893833
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 28816.38671875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 38.mlp.down_proj residual time: 1.1722891330718994
INFO - 38.mlp.down_proj 2ed gptq time: 9.778729677200317
INFO - 38.mlp.down_proj 2ed error time: 0.003107786178588867
INFO - 38.mlp.down_proj proxy error after residual VPTQ: 1.570057861499663e-07, 1.9424798665568233e-05, 0.00808274932205677
INFO - ----Quantizing layer 39 ...---- 2026-02-23 13:54:44 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/39_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:54:45 39.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:39.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 62 iterations, error 11981.1865234375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 39.self_attn.q_proj 1st kmeans time: 23.83582854270935
INFO - 39.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 39.self_attn.q_proj proxy error before VPTQ: 1.527730319139664e-06, 6.32564042462036e-05, 0.0241513941437006
INFO - 39.self_attn.q_proj 1st error time: 2.3109872341156006
INFO - 39.self_attn.q_proj proxy error after VPTQ: 9.63729917202727e-07, 6.32564042462036e-05, 0.0152352936565876
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 4828.33642578125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 39.self_attn.q_proj residual time: 0.31699490547180176
INFO - 39.self_attn.q_proj 2ed gptq time: 2.603982925415039
INFO - 39.self_attn.q_proj 2ed error time: 0.00039267539978027344
INFO - 39.self_attn.q_proj proxy error after residual VPTQ: 3.2850704201337066e-07, 6.32564042462036e-05, 0.005193261429667473
INFO - load Hessian from /workspace/hessians/quip-format/hessians/39_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:55:16 39.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:39.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 2526.3564453125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 39.self_attn.k_proj 1st kmeans time: 2.2805745601654053
INFO - 39.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 39.self_attn.k_proj proxy error before VPTQ: 3.021127668034751e-07, 1.4980223568272777e-05, 0.020167440176010132
INFO - 39.self_attn.k_proj 1st error time: 1.539219856262207
INFO - 39.self_attn.k_proj proxy error after VPTQ: 2.2766305107779772e-07, 1.4980223568272777e-05, 0.015197573229670525
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1070.354736328125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 39.self_attn.k_proj residual time: 0.09049677848815918
INFO - 39.self_attn.k_proj 2ed gptq time: 2.0469021797180176
INFO - 39.self_attn.k_proj 2ed error time: 0.0004360675811767578
INFO - 39.self_attn.k_proj proxy error after residual VPTQ: 7.591808781626241e-08, 1.4980223568272777e-05, 0.005067887715995312
INFO - load Hessian from /workspace/hessians/quip-format/hessians/39_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:55:23 39.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:39.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 35 iterations, error 2975.8466796875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 39.self_attn.v_proj 1st kmeans time: 3.5053281784057617
INFO - 39.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 39.self_attn.v_proj proxy error before VPTQ: 2.798841478579561e-07, 1.3064972335996572e-05, 0.021422483026981354
INFO - 39.self_attn.v_proj 1st error time: 1.538198471069336
INFO - 39.self_attn.v_proj proxy error after VPTQ: 2.0541564538234525e-07, 1.3064972335996572e-05, 0.015722623094916344
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1237.042724609375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 39.self_attn.v_proj residual time: 0.08961963653564453
INFO - 39.self_attn.v_proj 2ed gptq time: 2.0592904090881348
INFO - 39.self_attn.v_proj 2ed error time: 0.00046372413635253906
INFO - 39.self_attn.v_proj proxy error after residual VPTQ: 6.582374112440448e-08, 1.3064972335996572e-05, 0.005038184579461813
INFO - load Hessian from /workspace/hessians/quip-format/hessians/39_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:55:32 39.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:39.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 83 iterations, error 21913.994140625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 39.self_attn.o_proj 1st kmeans time: 31.977600812911987
INFO - 39.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 39.self_attn.o_proj proxy error before VPTQ: 2.150062755390536e-06, 5.948133912170306e-05, 0.036146845668554306
INFO - 39.self_attn.o_proj 1st error time: 2.3216872215270996
INFO - 39.self_attn.o_proj proxy error after VPTQ: 1.4451992456088192e-06, 5.948133912170306e-05, 0.024296682327985764
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 8208.7578125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 39.self_attn.o_proj residual time: 0.31217050552368164
INFO - 39.self_attn.o_proj 2ed gptq time: 2.6018807888031006
INFO - 39.self_attn.o_proj 2ed error time: 0.0011718273162841797
INFO - 39.self_attn.o_proj proxy error after residual VPTQ: 4.5598110887112853e-07, 5.948133912170306e-05, 0.007665952201932669
INFO - load Hessian from /workspace/hessians/quip-format/hessians/39_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:56:11 39.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:39.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 61835.5625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 39.mlp.gate_proj 1st kmeans time: 132.59569549560547
INFO - 39.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 39.mlp.gate_proj proxy error before VPTQ: 9.213181328959763e-06, 0.00018914253450930119, 0.048710256814956665
INFO - 39.mlp.gate_proj 1st error time: 4.760370492935181
INFO - 39.mlp.gate_proj proxy error after VPTQ: 6.556802418344887e-06, 0.00018914253450930119, 0.03466593474149704
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 22537.927734375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 39.mlp.gate_proj residual time: 1.056950330734253
INFO - 39.mlp.gate_proj 2ed gptq time: 5.05655837059021
INFO - 39.mlp.gate_proj 2ed error time: 0.0013341903686523438
INFO - 39.mlp.gate_proj proxy error after residual VPTQ: 2.1227458546491107e-06, 0.00018914253450930119, 0.011222995817661285
INFO - load Hessian from /workspace/hessians/quip-format/hessians/39_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 13:58:37 39.mlp.up_proj
INFO - enabling norm dim 0, layer_name:39.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 69757.3203125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 39.mlp.up_proj 1st kmeans time: 132.40331506729126
INFO - 39.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 39.mlp.up_proj proxy error before VPTQ: 7.745563380012754e-06, 0.000131763779791072, 0.05878370627760887
INFO - 39.mlp.up_proj 1st error time: 4.741165637969971
INFO - 39.mlp.up_proj proxy error after VPTQ: 5.431962108559674e-06, 0.000131763779791072, 0.04122500121593475
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 96 iterations, error 25182.80078125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 39.mlp.up_proj residual time: 1.0604102611541748
INFO - 39.mlp.up_proj 2ed gptq time: 5.131926536560059
INFO - 39.mlp.up_proj 2ed error time: 0.0014672279357910156
INFO - 39.mlp.up_proj proxy error after residual VPTQ: 1.744150040394743e-06, 0.000131763779791072, 0.013236946426331997
INFO - load Hessian from /workspace/hessians/quip-format/hessians/39_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:01:04 39.mlp.down_proj
INFO - enabling norm dim 0, layer_name:39.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 85549.984375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 39.mlp.down_proj 1st kmeans time: 132.47268986701965
INFO - 39.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 39.mlp.down_proj proxy error before VPTQ: 6.305770625658624e-07, 1.587070983077865e-05, 0.039732128381729126
INFO - 39.mlp.down_proj 1st error time: 8.568471431732178
INFO - 39.mlp.down_proj proxy error after VPTQ: 4.868912242272927e-07, 1.587070983077865e-05, 0.030678603798151016
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 29861.44921875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 39.mlp.down_proj residual time: 1.0763776302337646
INFO - 39.mlp.down_proj 2ed gptq time: 9.631439685821533
INFO - 39.mlp.down_proj 2ed error time: 0.0024945735931396484
INFO - 39.mlp.down_proj proxy error after residual VPTQ: 1.5858658741763065e-07, 1.587070983077865e-05, 0.009992406703531742
INFO - ----Quantizing layer 40 ...---- 2026-02-23 14:03:40 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/40_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:03:41 40.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:40.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 72 iterations, error 12337.9365234375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 40.self_attn.q_proj 1st kmeans time: 27.52479314804077
INFO - 40.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 40.self_attn.q_proj proxy error before VPTQ: 1.4684036386825028e-06, 5.34528044227045e-05, 0.0274710301309824
INFO - 40.self_attn.q_proj 1st error time: 2.3167479038238525
INFO - 40.self_attn.q_proj proxy error after VPTQ: 9.54566644395527e-07, 5.34528044227045e-05, 0.017858121544122696
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 4921.41064453125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 40.self_attn.q_proj residual time: 0.3145606517791748
INFO - 40.self_attn.q_proj 2ed gptq time: 2.652513027191162
INFO - 40.self_attn.q_proj 2ed error time: 0.0004024505615234375
INFO - 40.self_attn.q_proj proxy error after residual VPTQ: 3.2447093190057785e-07, 5.34528044227045e-05, 0.006070232018828392
INFO - load Hessian from /workspace/hessians/quip-format/hessians/40_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:04:16 40.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:40.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 33 iterations, error 2578.3876953125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 40.self_attn.k_proj 1st kmeans time: 3.2934186458587646
INFO - 40.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 40.self_attn.k_proj proxy error before VPTQ: 3.0113594107206154e-07, 1.478131798648974e-05, 0.02037273906171322
INFO - 40.self_attn.k_proj 1st error time: 1.5475409030914307
INFO - 40.self_attn.k_proj proxy error after VPTQ: 2.2659827436655178e-07, 1.478131798648974e-05, 0.015330045484006405
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1091.4169921875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 40.self_attn.k_proj residual time: 0.0905914306640625
INFO - 40.self_attn.k_proj 2ed gptq time: 2.0597827434539795
INFO - 40.self_attn.k_proj 2ed error time: 0.0005140304565429688
INFO - 40.self_attn.k_proj proxy error after residual VPTQ: 7.614173824777026e-08, 1.478131798648974e-05, 0.0051512145437300205
INFO - load Hessian from /workspace/hessians/quip-format/hessians/40_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:04:24 40.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:40.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 23 iterations, error 3455.31005859375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 40.self_attn.v_proj 1st kmeans time: 2.3589797019958496
INFO - 40.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 40.self_attn.v_proj proxy error before VPTQ: 3.367380827512534e-07, 1.7911333998199552e-05, 0.01880027912557125
INFO - 40.self_attn.v_proj 1st error time: 1.5432884693145752
INFO - 40.self_attn.v_proj proxy error after VPTQ: 2.2798758436692879e-07, 1.7911333998199552e-05, 0.0127286771312356
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1362.6468505859375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 40.self_attn.v_proj residual time: 0.09151935577392578
INFO - 40.self_attn.v_proj 2ed gptq time: 2.064925193786621
INFO - 40.self_attn.v_proj 2ed error time: 0.0004856586456298828
INFO - 40.self_attn.v_proj proxy error after residual VPTQ: 7.355474451742339e-08, 1.7911333998199552e-05, 0.004106603562831879
INFO - load Hessian from /workspace/hessians/quip-format/hessians/40_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:04:32 40.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:40.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 81 iterations, error 20574.41796875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 40.self_attn.o_proj 1st kmeans time: 30.877047777175903
INFO - 40.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 40.self_attn.o_proj proxy error before VPTQ: 2.239229615952354e-06, 5.170582880964503e-05, 0.0433071032166481
INFO - 40.self_attn.o_proj 1st error time: 2.319148302078247
INFO - 40.self_attn.o_proj proxy error after VPTQ: 1.4886966255289735e-06, 5.170582880964503e-05, 0.02879166044294834
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7715.35205078125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 40.self_attn.o_proj residual time: 0.31647801399230957
INFO - 40.self_attn.o_proj 2ed gptq time: 2.6052515506744385
INFO - 40.self_attn.o_proj 2ed error time: 0.0011391639709472656
INFO - 40.self_attn.o_proj proxy error after residual VPTQ: 4.775620823238569e-07, 5.170582880964503e-05, 0.009236136451363564
INFO - load Hessian from /workspace/hessians/quip-format/hessians/40_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:05:10 40.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:40.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 65089.98046875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 40.mlp.gate_proj 1st kmeans time: 132.50353264808655
INFO - 40.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 40.mlp.gate_proj proxy error before VPTQ: 9.09118534764275e-06, 0.00016992897144518793, 0.053499914705753326
INFO - 40.mlp.gate_proj 1st error time: 4.765315294265747
INFO - 40.mlp.gate_proj proxy error after VPTQ: 6.559586381627014e-06, 0.00016992897144518793, 0.03860193118453026
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 90 iterations, error 23494.357421875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 40.mlp.gate_proj residual time: 0.9513387680053711
INFO - 40.mlp.gate_proj 2ed gptq time: 5.055721282958984
INFO - 40.mlp.gate_proj 2ed error time: 0.0013172626495361328
INFO - 40.mlp.gate_proj proxy error after residual VPTQ: 2.125955688825343e-06, 0.00016992897144518793, 0.01251084916293621
INFO - load Hessian from /workspace/hessians/quip-format/hessians/40_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:07:35 40.mlp.up_proj
INFO - enabling norm dim 0, layer_name:40.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 71376.84375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 40.mlp.up_proj 1st kmeans time: 132.400146484375
INFO - 40.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 40.mlp.up_proj proxy error before VPTQ: 7.75786247686483e-06, 0.00012780161341652274, 0.060702383518218994
INFO - 40.mlp.up_proj 1st error time: 4.745440721511841
INFO - 40.mlp.up_proj proxy error after VPTQ: 5.516325927601429e-06, 0.00012780161341652274, 0.04316319525241852
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 25558.91796875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 40.mlp.up_proj residual time: 1.046971082687378
INFO - 40.mlp.up_proj 2ed gptq time: 5.131486177444458
INFO - 40.mlp.up_proj 2ed error time: 0.0011179447174072266
INFO - 40.mlp.up_proj proxy error after residual VPTQ: 1.7681238659861265e-06, 0.00012780161341652274, 0.013834910467267036
INFO - load Hessian from /workspace/hessians/quip-format/hessians/40_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:10:03 40.mlp.down_proj
INFO - enabling norm dim 0, layer_name:40.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 86528.8515625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 40.mlp.down_proj 1st kmeans time: 132.75051951408386
INFO - 40.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 40.mlp.down_proj proxy error before VPTQ: 6.387112989614252e-07, 1.679592241998762e-05, 0.03802775964140892
INFO - 40.mlp.down_proj 1st error time: 8.548003435134888
INFO - 40.mlp.down_proj proxy error after VPTQ: 5.10223571836832e-07, 1.679592241998762e-05, 0.0303778238594532
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 30043.865234375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 40.mlp.down_proj residual time: 1.0075314044952393
INFO - 40.mlp.down_proj 2ed gptq time: 9.645038843154907
INFO - 40.mlp.down_proj 2ed error time: 0.002374410629272461
INFO - 40.mlp.down_proj proxy error after residual VPTQ: 1.6632415622552799e-07, 1.679592241998762e-05, 0.009902650490403175
INFO - ----Quantizing layer 41 ...---- 2026-02-23 14:12:39 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/41_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:12:39 41.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:41.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 83 iterations, error 14041.251953125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 41.self_attn.q_proj 1st kmeans time: 31.59596037864685
INFO - 41.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 41.self_attn.q_proj proxy error before VPTQ: 1.5769045376146096e-06, 8.453273039776832e-05, 0.018654366955161095
INFO - 41.self_attn.q_proj 1st error time: 2.305753469467163
INFO - 41.self_attn.q_proj proxy error after VPTQ: 1.0337555522710318e-06, 8.453273039776832e-05, 0.012229057028889656
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 5526.51171875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 41.self_attn.q_proj residual time: 0.31696319580078125
INFO - 41.self_attn.q_proj 2ed gptq time: 2.6043643951416016
INFO - 41.self_attn.q_proj 2ed error time: 0.0003635883331298828
INFO - 41.self_attn.q_proj proxy error after residual VPTQ: 3.465599434093747e-07, 8.453273039776832e-05, 0.004099713172763586
INFO - load Hessian from /workspace/hessians/quip-format/hessians/41_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:13:18 41.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:41.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 2902.291015625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 41.self_attn.k_proj 1st kmeans time: 2.2635631561279297
INFO - 41.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 41.self_attn.k_proj proxy error before VPTQ: 3.2862510579434456e-07, 1.5252506273100153e-05, 0.021545646712183952
INFO - 41.self_attn.k_proj 1st error time: 1.546475887298584
INFO - 41.self_attn.k_proj proxy error after VPTQ: 2.5519540258756024e-07, 1.5252506273100153e-05, 0.016731375828385353
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1202.232421875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 41.self_attn.k_proj residual time: 0.09079098701477051
INFO - 41.self_attn.k_proj 2ed gptq time: 2.0568034648895264
INFO - 41.self_attn.k_proj 2ed error time: 0.0004215240478515625
INFO - 41.self_attn.k_proj proxy error after residual VPTQ: 8.437536536121115e-08, 1.5252506273100153e-05, 0.005531901493668556
INFO - load Hessian from /workspace/hessians/quip-format/hessians/41_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:13:25 41.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:41.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 25 iterations, error 3365.728759765625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 41.self_attn.v_proj 1st kmeans time: 2.5509684085845947
INFO - 41.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 41.self_attn.v_proj proxy error before VPTQ: 3.4477062627047417e-07, 1.0680583727662452e-05, 0.032280128449201584
INFO - 41.self_attn.v_proj 1st error time: 1.5414559841156006
INFO - 41.self_attn.v_proj proxy error after VPTQ: 2.6239592898491537e-07, 1.0680583727662452e-05, 0.024567564949393272
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1351.728515625
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 41.self_attn.v_proj residual time: 0.09077811241149902
INFO - 41.self_attn.v_proj 2ed gptq time: 2.0630829334259033
INFO - 41.self_attn.v_proj 2ed error time: 0.00043845176696777344
INFO - 41.self_attn.v_proj proxy error after residual VPTQ: 8.434922449396254e-08, 1.0680583727662452e-05, 0.007897435687482357
INFO - load Hessian from /workspace/hessians/quip-format/hessians/41_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:13:33 41.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:41.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 76 iterations, error 21430.146484375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 41.self_attn.o_proj 1st kmeans time: 29.03388285636902
INFO - 41.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 41.self_attn.o_proj proxy error before VPTQ: 2.6562406674202066e-06, 0.00012919607979711145, 0.02055976167321205
INFO - 41.self_attn.o_proj 1st error time: 2.318631887435913
INFO - 41.self_attn.o_proj proxy error after VPTQ: 1.72606814885512e-06, 0.00012919607979711145, 0.013360066339373589
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 96 iterations, error 8091.0537109375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 41.self_attn.o_proj residual time: 0.30556535720825195
INFO - 41.self_attn.o_proj 2ed gptq time: 2.678431987762451
INFO - 41.self_attn.o_proj 2ed error time: 0.0010900497436523438
INFO - 41.self_attn.o_proj proxy error after residual VPTQ: 5.475528723764e-07, 0.00012919607979711145, 0.00423815380781889
INFO - load Hessian from /workspace/hessians/quip-format/hessians/41_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:14:09 41.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:41.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 67388.5078125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 41.mlp.gate_proj 1st kmeans time: 132.41363859176636
INFO - 41.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 41.mlp.gate_proj proxy error before VPTQ: 9.107488949666731e-06, 0.00018652481958270073, 0.04882722347974777
INFO - 41.mlp.gate_proj 1st error time: 4.957274675369263
INFO - 41.mlp.gate_proj proxy error after VPTQ: 6.648114322160836e-06, 0.00018652481958270073, 0.03564198315143585
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 24245.05859375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 41.mlp.gate_proj residual time: 1.047837495803833
INFO - 41.mlp.gate_proj 2ed gptq time: 5.057679176330566
INFO - 41.mlp.gate_proj 2ed error time: 0.0013887882232666016
INFO - 41.mlp.gate_proj proxy error after residual VPTQ: 2.1429350454127416e-06, 0.00018652481958270073, 0.011488739401102066
INFO - load Hessian from /workspace/hessians/quip-format/hessians/41_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:16:35 41.mlp.up_proj
INFO - enabling norm dim 0, layer_name:41.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 72261.65625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 41.mlp.up_proj 1st kmeans time: 132.4122576713562
INFO - 41.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 41.mlp.up_proj proxy error before VPTQ: 7.878067663114052e-06, 0.00012223377416376024, 0.06445082277059555
INFO - 41.mlp.up_proj 1st error time: 4.812875509262085
INFO - 41.mlp.up_proj proxy error after VPTQ: 5.706389856641181e-06, 0.00012223377416376024, 0.046684231609106064
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 99 iterations, error 25763.3515625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 41.mlp.up_proj residual time: 1.0822408199310303
INFO - 41.mlp.up_proj 2ed gptq time: 5.080367088317871
INFO - 41.mlp.up_proj 2ed error time: 0.0012416839599609375
INFO - 41.mlp.up_proj proxy error after residual VPTQ: 1.8292573713551974e-06, 0.00012223377416376024, 0.014965237118303776
INFO - load Hessian from /workspace/hessians/quip-format/hessians/41_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:19:02 41.mlp.down_proj
INFO - enabling norm dim 0, layer_name:41.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 86589.2421875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 41.mlp.down_proj 1st kmeans time: 132.392564535141
INFO - 41.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 41.mlp.down_proj proxy error before VPTQ: 6.20792434347095e-07, 1.1705339602485765e-05, 0.05303497985005379
INFO - 41.mlp.down_proj 1st error time: 8.930055618286133
INFO - 41.mlp.down_proj proxy error after VPTQ: 5.098727910990419e-07, 1.1705339602485765e-05, 0.0435589924454689
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 29930.41015625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 41.mlp.down_proj residual time: 1.10581636428833
INFO - 41.mlp.down_proj 2ed gptq time: 9.708519458770752
INFO - 41.mlp.down_proj 2ed error time: 0.03540301322937012
INFO - 41.mlp.down_proj proxy error after residual VPTQ: 1.66344165108967e-07, 1.1705339602485765e-05, 0.014210964553058147
INFO - ----Quantizing layer 42 ...---- 2026-02-23 14:21:40 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/42_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:21:41 42.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:42.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 65 iterations, error 13819.8681640625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 42.self_attn.q_proj 1st kmeans time: 24.94537115097046
INFO - 42.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 42.self_attn.q_proj proxy error before VPTQ: 1.5618720681231935e-06, 5.182152381166816e-05, 0.030139446258544922
INFO - 42.self_attn.q_proj 1st error time: 2.2723748683929443
INFO - 42.self_attn.q_proj proxy error after VPTQ: 1.0881412890739739e-06, 5.182152381166816e-05, 0.02099786326289177
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 5408.35595703125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 42.self_attn.q_proj residual time: 0.31444454193115234
INFO - 42.self_attn.q_proj 2ed gptq time: 2.5724847316741943
INFO - 42.self_attn.q_proj 2ed error time: 0.0002903938293457031
INFO - 42.self_attn.q_proj proxy error after residual VPTQ: 3.662420340333483e-07, 5.182152381166816e-05, 0.007067373022437096
INFO - load Hessian from /workspace/hessians/quip-format/hessians/42_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:22:12 42.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:42.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 27 iterations, error 3124.08935546875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 42.self_attn.k_proj 1st kmeans time: 2.7330851554870605
INFO - 42.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 42.self_attn.k_proj proxy error before VPTQ: 3.2575525210631895e-07, 1.2423196494637523e-05, 0.026221532374620438
INFO - 42.self_attn.k_proj 1st error time: 1.501326322555542
INFO - 42.self_attn.k_proj proxy error after VPTQ: 2.5528336777824734e-07, 1.2423196494637523e-05, 0.020548928529024124
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1273.7333984375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 42.self_attn.k_proj residual time: 0.09465956687927246
INFO - 42.self_attn.k_proj 2ed gptq time: 2.0615808963775635
INFO - 42.self_attn.k_proj 2ed error time: 0.0007302761077880859
INFO - 42.self_attn.k_proj proxy error after residual VPTQ: 8.440990484359645e-08, 1.2423196494637523e-05, 0.006794539745897055
INFO - load Hessian from /workspace/hessians/quip-format/hessians/42_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:22:20 42.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:42.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 3553.537109375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 42.self_attn.v_proj 1st kmeans time: 2.2769033908843994
INFO - 42.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 42.self_attn.v_proj proxy error before VPTQ: 3.711748775003798e-07, 1.5946778148645535e-05, 0.02327585406601429
INFO - 42.self_attn.v_proj 1st error time: 1.4995677471160889
INFO - 42.self_attn.v_proj proxy error after VPTQ: 2.885915932893113e-07, 1.5946778148645535e-05, 0.018097171559929848
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1412.53466796875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 42.self_attn.v_proj residual time: 0.09307575225830078
INFO - 42.self_attn.v_proj 2ed gptq time: 1.9972357749938965
INFO - 42.self_attn.v_proj 2ed error time: 0.0004820823669433594
INFO - 42.self_attn.v_proj proxy error after residual VPTQ: 9.257982469534909e-08, 1.5946778148645535e-05, 0.00580555060878396
INFO - load Hessian from /workspace/hessians/quip-format/hessians/42_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:22:28 42.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:42.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 62 iterations, error 21289.125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 42.self_attn.o_proj 1st kmeans time: 23.82589554786682
INFO - 42.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 42.self_attn.o_proj proxy error before VPTQ: 2.4385308279306628e-06, 5.003308615414426e-05, 0.04873836413025856
INFO - 42.self_attn.o_proj 1st error time: 2.3234736919403076
INFO - 42.self_attn.o_proj proxy error after VPTQ: 1.6867174963408615e-06, 5.003308615414426e-05, 0.03371204063296318
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 8046.3310546875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 42.self_attn.o_proj residual time: 0.31041669845581055
INFO - 42.self_attn.o_proj 2ed gptq time: 2.6103739738464355
INFO - 42.self_attn.o_proj 2ed error time: 0.0017046928405761719
INFO - 42.self_attn.o_proj proxy error after residual VPTQ: 5.396300934989995e-07, 5.003308615414426e-05, 0.010785465128719807
INFO - load Hessian from /workspace/hessians/quip-format/hessians/42_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:22:58 42.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:42.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 68340.3125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 42.mlp.gate_proj 1st kmeans time: 132.4078974723816
INFO - 42.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 42.mlp.gate_proj proxy error before VPTQ: 9.229188435710967e-06, 0.0002062538405880332, 0.04474674537777901
INFO - 42.mlp.gate_proj 1st error time: 4.844468593597412
INFO - 42.mlp.gate_proj proxy error after VPTQ: 6.830785423517227e-06, 0.0002062538405880332, 0.033118341118097305
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 95 iterations, error 24359.203125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 42.mlp.gate_proj residual time: 1.0480830669403076
INFO - 42.mlp.gate_proj 2ed gptq time: 5.056445121765137
INFO - 42.mlp.gate_proj 2ed error time: 0.0013508796691894531
INFO - 42.mlp.gate_proj proxy error after residual VPTQ: 2.204043084930163e-06, 0.0002062538405880332, 0.010686070658266544
INFO - load Hessian from /workspace/hessians/quip-format/hessians/42_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:25:24 42.mlp.up_proj
INFO - enabling norm dim 0, layer_name:42.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 72890.53125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 42.mlp.up_proj 1st kmeans time: 132.393803358078
INFO - 42.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 42.mlp.up_proj proxy error before VPTQ: 8.01172063802369e-06, 0.00013606670836452395, 0.05888082832098007
INFO - 42.mlp.up_proj 1st error time: 4.726155042648315
INFO - 42.mlp.up_proj proxy error after VPTQ: 5.937165042269044e-06, 0.00013606670836452395, 0.04363422095775604
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 95 iterations, error 25891.365234375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 42.mlp.up_proj residual time: 0.9685583114624023
INFO - 42.mlp.up_proj 2ed gptq time: 5.017575263977051
INFO - 42.mlp.up_proj 2ed error time: 0.0011594295501708984
INFO - 42.mlp.up_proj proxy error after residual VPTQ: 1.903933707581018e-06, 0.00013606670836452395, 0.013992649503052235
INFO - load Hessian from /workspace/hessians/quip-format/hessians/42_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:27:51 42.mlp.down_proj
INFO - enabling norm dim 0, layer_name:42.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 85754.75
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 42.mlp.down_proj 1st kmeans time: 132.45742893218994
INFO - 42.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 42.mlp.down_proj proxy error before VPTQ: 6.146452165012306e-07, 1.1556405297596939e-05, 0.05318653956055641
INFO - 42.mlp.down_proj 1st error time: 8.575756311416626
INFO - 42.mlp.down_proj proxy error after VPTQ: 5.088339207759418e-07, 1.1556405297596939e-05, 0.04403046518564224
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 29701.841796875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 42.mlp.down_proj residual time: 1.0756258964538574
INFO - 42.mlp.down_proj 2ed gptq time: 9.632440090179443
INFO - 42.mlp.down_proj 2ed error time: 0.002447366714477539
INFO - 42.mlp.down_proj proxy error after residual VPTQ: 1.6636126076718938e-07, 1.1556405297596939e-05, 0.014395589008927345
INFO - ----Quantizing layer 43 ...---- 2026-02-23 14:30:28 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/43_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:30:28 43.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:43.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 78 iterations, error 14058.048828125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 43.self_attn.q_proj 1st kmeans time: 29.744844436645508
INFO - 43.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 43.self_attn.q_proj proxy error before VPTQ: 1.5343648556154221e-06, 5.4362506489269435e-05, 0.02822468988597393
INFO - 43.self_attn.q_proj 1st error time: 2.3078436851501465
INFO - 43.self_attn.q_proj proxy error after VPTQ: 1.0786707207444124e-06, 5.4362506489269435e-05, 0.019842181354761124
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 5476.1220703125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 43.self_attn.q_proj residual time: 0.3153684139251709
INFO - 43.self_attn.q_proj 2ed gptq time: 2.595978021621704
INFO - 43.self_attn.q_proj 2ed error time: 0.0004112720489501953
INFO - 43.self_attn.q_proj proxy error after residual VPTQ: 3.638018029050727e-07, 5.4362506489269435e-05, 0.006692145485430956
INFO - load Hessian from /workspace/hessians/quip-format/hessians/43_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:31:05 43.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:43.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 26 iterations, error 3216.73583984375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 43.self_attn.k_proj 1st kmeans time: 2.64081072807312
INFO - 43.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 43.self_attn.k_proj proxy error before VPTQ: 3.080565420532366e-07, 1.1749294571927749e-05, 0.02621915191411972
INFO - 43.self_attn.k_proj 1st error time: 1.5075485706329346
INFO - 43.self_attn.k_proj proxy error after VPTQ: 2.4404957343904243e-07, 1.1749294571927749e-05, 0.02077142335474491
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1308.329833984375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 43.self_attn.k_proj residual time: 0.09132266044616699
INFO - 43.self_attn.k_proj 2ed gptq time: 1.9567580223083496
INFO - 43.self_attn.k_proj 2ed error time: 0.00037741661071777344
INFO - 43.self_attn.k_proj proxy error after residual VPTQ: 8.081670443971234e-08, 1.1749294571927749e-05, 0.006878430489450693
INFO - load Hessian from /workspace/hessians/quip-format/hessians/43_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:31:13 43.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:43.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 3632.639404296875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 43.self_attn.v_proj 1st kmeans time: 2.2653584480285645
INFO - 43.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 43.self_attn.v_proj proxy error before VPTQ: 4.1778784520829504e-07, 1.0768168976937886e-05, 0.038798410445451736
INFO - 43.self_attn.v_proj 1st error time: 1.5416395664215088
INFO - 43.self_attn.v_proj proxy error after VPTQ: 3.2928807058851817e-07, 1.0768168976937886e-05, 0.030579764395952225
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1440.830810546875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 43.self_attn.v_proj residual time: 0.0984044075012207
INFO - 43.self_attn.v_proj 2ed gptq time: 2.0454177856445312
INFO - 43.self_attn.v_proj 2ed error time: 0.0004591941833496094
INFO - 43.self_attn.v_proj proxy error after residual VPTQ: 1.0604175315620523e-07, 1.0768168976937886e-05, 0.009847705252468586
INFO - load Hessian from /workspace/hessians/quip-format/hessians/43_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:31:20 43.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:43.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 82 iterations, error 22137.916015625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 43.self_attn.o_proj 1st kmeans time: 31.49140977859497
INFO - 43.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 43.self_attn.o_proj proxy error before VPTQ: 2.4785886125755496e-06, 6.082266918383539e-05, 0.04075106605887413
INFO - 43.self_attn.o_proj 1st error time: 2.38554048538208
INFO - 43.self_attn.o_proj proxy error after VPTQ: 1.6925079080465366e-06, 6.082266918383539e-05, 0.027826925739645958
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 8130.1220703125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 43.self_attn.o_proj residual time: 0.3150794506072998
INFO - 43.self_attn.o_proj 2ed gptq time: 2.6106419563293457
INFO - 43.self_attn.o_proj 2ed error time: 0.0011446475982666016
INFO - 43.self_attn.o_proj proxy error after residual VPTQ: 5.374286047299393e-07, 6.082266918383539e-05, 0.008835991844534874
INFO - load Hessian from /workspace/hessians/quip-format/hessians/43_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:31:59 43.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:43.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 68754.4453125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 43.mlp.gate_proj 1st kmeans time: 132.51223707199097
INFO - 43.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 43.mlp.gate_proj proxy error before VPTQ: 9.550996765028685e-06, 0.00020387076074257493, 0.04684828966856003
INFO - 43.mlp.gate_proj 1st error time: 4.972446918487549
INFO - 43.mlp.gate_proj proxy error after VPTQ: 7.217392521852162e-06, 0.00020387076074257493, 0.03540180251002312
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 97 iterations, error 24441.90234375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 43.mlp.gate_proj residual time: 1.0376744270324707
INFO - 43.mlp.gate_proj 2ed gptq time: 5.0434300899505615
INFO - 43.mlp.gate_proj 2ed error time: 0.0013968944549560547
INFO - 43.mlp.gate_proj proxy error after residual VPTQ: 2.3363143100141315e-06, 0.00020387076074257493, 0.011459780856966972
INFO - load Hessian from /workspace/hessians/quip-format/hessians/43_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:34:25 43.mlp.up_proj
INFO - enabling norm dim 0, layer_name:43.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 73303.734375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 43.mlp.up_proj 1st kmeans time: 132.41124963760376
INFO - 43.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 43.mlp.up_proj proxy error before VPTQ: 8.024688213481568e-06, 0.0001293353852815926, 0.06204557418823242
INFO - 43.mlp.up_proj 1st error time: 4.752316951751709
INFO - 43.mlp.up_proj proxy error after VPTQ: 5.947059435129631e-06, 0.0001293353852815926, 0.04598169028759003
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 25886.189453125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 43.mlp.up_proj residual time: 1.0363810062408447
INFO - 43.mlp.up_proj 2ed gptq time: 5.07478928565979
INFO - 43.mlp.up_proj 2ed error time: 0.0014450550079345703
INFO - 43.mlp.up_proj proxy error after residual VPTQ: 1.9088165572611615e-06, 0.0001293353852815926, 0.014758656732738018
INFO - load Hessian from /workspace/hessians/quip-format/hessians/43_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:36:52 43.mlp.down_proj
INFO - enabling norm dim 0, layer_name:43.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 84167.7890625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 43.mlp.down_proj 1st kmeans time: 132.53389930725098
INFO - 43.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 43.mlp.down_proj proxy error before VPTQ: 6.119926752035099e-07, 1.2089293704775628e-05, 0.05062269791960716
INFO - 43.mlp.down_proj 1st error time: 8.57975697517395
INFO - 43.mlp.down_proj proxy error after VPTQ: 5.080872824692051e-07, 1.2089293704775628e-05, 0.04202787205576897
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 29156.654296875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 43.mlp.down_proj residual time: 1.0640957355499268
INFO - 43.mlp.down_proj 2ed gptq time: 9.46730375289917
INFO - 43.mlp.down_proj 2ed error time: 0.002531766891479492
INFO - 43.mlp.down_proj proxy error after residual VPTQ: 1.6645337552745332e-07, 1.2089293704775628e-05, 0.01376865990459919
INFO - ----Quantizing layer 44 ...---- 2026-02-23 14:39:29 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/44_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:39:29 44.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:44.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 79 iterations, error 14014.8037109375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 44.self_attn.q_proj 1st kmeans time: 30.135842084884644
INFO - 44.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 44.self_attn.q_proj proxy error before VPTQ: 1.5094931313797133e-06, 4.678584809880704e-05, 0.03226388245820999
INFO - 44.self_attn.q_proj 1st error time: 2.253948450088501
INFO - 44.self_attn.q_proj proxy error after VPTQ: 1.0695907803892624e-06, 4.678584809880704e-05, 0.02286141738295555
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 5456.73388671875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 44.self_attn.q_proj residual time: 0.3563215732574463
INFO - 44.self_attn.q_proj 2ed gptq time: 2.5480268001556396
INFO - 44.self_attn.q_proj 2ed error time: 0.0002713203430175781
INFO - 44.self_attn.q_proj proxy error after residual VPTQ: 3.6184593454891e-07, 4.678584809880704e-05, 0.007734089624136686
INFO - load Hessian from /workspace/hessians/quip-format/hessians/44_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:40:06 44.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:44.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 28 iterations, error 3202.79443359375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 44.self_attn.k_proj 1st kmeans time: 2.848378896713257
INFO - 44.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 44.self_attn.k_proj proxy error before VPTQ: 2.976873361149046e-07, 1.1972313586738892e-05, 0.024864645674824715
INFO - 44.self_attn.k_proj 1st error time: 1.526947021484375
INFO - 44.self_attn.k_proj proxy error after VPTQ: 2.3814844496428123e-07, 1.1972313586738892e-05, 0.01989159733057022
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1301.2777099609375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 44.self_attn.k_proj residual time: 0.09796333312988281
INFO - 44.self_attn.k_proj 2ed gptq time: 1.9442172050476074
INFO - 44.self_attn.k_proj 2ed error time: 0.00041866302490234375
INFO - 44.self_attn.k_proj proxy error after residual VPTQ: 7.942534807625634e-08, 1.1972313586738892e-05, 0.006634084973484278
INFO - load Hessian from /workspace/hessians/quip-format/hessians/44_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:40:14 44.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:44.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 26 iterations, error 3616.87109375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 44.self_attn.v_proj 1st kmeans time: 2.6555614471435547
INFO - 44.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 44.self_attn.v_proj proxy error before VPTQ: 4.305835545892478e-07, 1.2244205208844505e-05, 0.035166312009096146
INFO - 44.self_attn.v_proj 1st error time: 1.518855333328247
INFO - 44.self_attn.v_proj proxy error after VPTQ: 3.439799058924109e-07, 1.2244205208844505e-05, 0.028093282133340836
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1427.65380859375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 44.self_attn.v_proj residual time: 0.09393572807312012
INFO - 44.self_attn.v_proj 2ed gptq time: 1.9809517860412598
INFO - 44.self_attn.v_proj 2ed error time: 0.0003676414489746094
INFO - 44.self_attn.v_proj proxy error after residual VPTQ: 1.1028046031924532e-07, 1.2244205208844505e-05, 0.009006747044622898
INFO - load Hessian from /workspace/hessians/quip-format/hessians/44_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:40:22 44.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:44.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 64 iterations, error 21845.0
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 44.self_attn.o_proj 1st kmeans time: 24.78711199760437
INFO - 44.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 44.self_attn.o_proj proxy error before VPTQ: 2.483613116055494e-06, 9.454668906982988e-05, 0.026268642395734787
INFO - 44.self_attn.o_proj 1st error time: 2.2780590057373047
INFO - 44.self_attn.o_proj proxy error after VPTQ: 1.7222690757989767e-06, 9.454668906982988e-05, 0.018216069787740707
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 8037.8125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 44.self_attn.o_proj residual time: 0.3083992004394531
INFO - 44.self_attn.o_proj 2ed gptq time: 2.574427843093872
INFO - 44.self_attn.o_proj 2ed error time: 0.0010066032409667969
INFO - 44.self_attn.o_proj proxy error after residual VPTQ: 5.47825266039581e-07, 9.454668906982988e-05, 0.005794229917228222
INFO - load Hessian from /workspace/hessians/quip-format/hessians/44_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:40:54 44.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:44.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 69173.65625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 44.mlp.gate_proj 1st kmeans time: 132.8977334499359
INFO - 44.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 44.mlp.gate_proj proxy error before VPTQ: 9.305391358793713e-06, 0.0002042931882897392, 0.04554919898509979
INFO - 44.mlp.gate_proj 1st error time: 4.729225158691406
INFO - 44.mlp.gate_proj proxy error after VPTQ: 6.998865501373075e-06, 0.0002042931882897392, 0.03425892814993858
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 90 iterations, error 24441.318359375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 44.mlp.gate_proj residual time: 0.9753284454345703
INFO - 44.mlp.gate_proj 2ed gptq time: 5.028545141220093
INFO - 44.mlp.gate_proj 2ed error time: 0.0011878013610839844
INFO - 44.mlp.gate_proj proxy error after residual VPTQ: 2.2579181404580595e-06, 0.0002042931882897392, 0.011052342131733894
INFO - load Hessian from /workspace/hessians/quip-format/hessians/44_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:43:20 44.mlp.up_proj
INFO - enabling norm dim 0, layer_name:44.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 73101.921875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 44.mlp.up_proj 1st kmeans time: 132.50614142417908
INFO - 44.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 44.mlp.up_proj proxy error before VPTQ: 8.257318768301047e-06, 0.00013384975318331271, 0.06169095262885094
INFO - 44.mlp.up_proj 1st error time: 4.804309368133545
INFO - 44.mlp.up_proj proxy error after VPTQ: 6.211213076312561e-06, 0.00013384975318331271, 0.046404365450143814
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 92 iterations, error 25763.005859375
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 44.mlp.up_proj residual time: 0.9872150421142578
INFO - 44.mlp.up_proj 2ed gptq time: 5.028447151184082
INFO - 44.mlp.up_proj 2ed error time: 0.0011954307556152344
INFO - 44.mlp.up_proj proxy error after residual VPTQ: 1.994534841287532e-06, 0.00013384975318331271, 0.014901296235620975
INFO - load Hessian from /workspace/hessians/quip-format/hessians/44_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:45:47 44.mlp.down_proj
INFO - enabling norm dim 0, layer_name:44.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 84119.609375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 44.mlp.down_proj 1st kmeans time: 132.64047145843506
INFO - 44.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 44.mlp.down_proj proxy error before VPTQ: 6.149747378003667e-07, 1.0974992619594559e-05, 0.05603418126702309
INFO - 44.mlp.down_proj 1st error time: 8.746630907058716
INFO - 44.mlp.down_proj proxy error after VPTQ: 5.132677074470848e-07, 1.0974992619594559e-05, 0.04676702246069908
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 29190.337890625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 44.mlp.down_proj residual time: 1.0175151824951172
INFO - 44.mlp.down_proj 2ed gptq time: 9.507889986038208
INFO - 44.mlp.down_proj 2ed error time: 0.0032312870025634766
INFO - 44.mlp.down_proj proxy error after residual VPTQ: 1.6838647809436225e-07, 1.0974992619594559e-05, 0.015342742204666138
INFO - ----Quantizing layer 45 ...---- 2026-02-23 14:48:24 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/45_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:48:25 45.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:45.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 73 iterations, error 15413.1650390625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 45.self_attn.q_proj 1st kmeans time: 27.8847336769104
INFO - 45.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 45.self_attn.q_proj proxy error before VPTQ: 1.6928636341617675e-06, 5.833629620610736e-05, 0.02901904471218586
INFO - 45.self_attn.q_proj 1st error time: 2.262324333190918
INFO - 45.self_attn.q_proj proxy error after VPTQ: 1.2112554941268172e-06, 5.833629620610736e-05, 0.020763324573636055
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 5916.830078125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 45.self_attn.q_proj residual time: 0.3123478889465332
INFO - 45.self_attn.q_proj 2ed gptq time: 2.5582447052001953
INFO - 45.self_attn.q_proj 2ed error time: 0.0003185272216796875
INFO - 45.self_attn.q_proj proxy error after residual VPTQ: 4.0471746842740686e-07, 5.833629620610736e-05, 0.006937661208212376
INFO - load Hessian from /workspace/hessians/quip-format/hessians/45_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:48:59 45.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:45.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 3633.36572265625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 45.self_attn.k_proj 1st kmeans time: 2.2669801712036133
INFO - 45.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 45.self_attn.k_proj proxy error before VPTQ: 3.7085681015014416e-07, 1.5770716345286928e-05, 0.02351553365588188
INFO - 45.self_attn.k_proj 1st error time: 1.497312068939209
INFO - 45.self_attn.k_proj proxy error after VPTQ: 2.923612782979035e-07, 1.5770716345286928e-05, 0.018538236618041992
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1464.0400390625
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 45.self_attn.k_proj residual time: 0.0973958969116211
INFO - 45.self_attn.k_proj 2ed gptq time: 2.0288195610046387
INFO - 45.self_attn.k_proj 2ed error time: 0.0004379749298095703
INFO - 45.self_attn.k_proj proxy error after residual VPTQ: 9.597484051937499e-08, 1.5770716345286928e-05, 0.006085636094212532
INFO - load Hessian from /workspace/hessians/quip-format/hessians/45_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:49:07 45.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:45.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 26 iterations, error 4069.6455078125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 45.self_attn.v_proj 1st kmeans time: 2.650876045227051
INFO - 45.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 45.self_attn.v_proj proxy error before VPTQ: 4.6899180006221286e-07, 1.4848321370664053e-05, 0.031585510820150375
INFO - 45.self_attn.v_proj 1st error time: 1.4974102973937988
INFO - 45.self_attn.v_proj proxy error after VPTQ: 3.6910844869453285e-07, 1.4848321370664053e-05, 0.024858597666025162
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1614.8482666015625
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 45.self_attn.v_proj residual time: 0.09538102149963379
INFO - 45.self_attn.v_proj 2ed gptq time: 1.972419023513794
INFO - 45.self_attn.v_proj 2ed error time: 0.0004150867462158203
INFO - 45.self_attn.v_proj proxy error after residual VPTQ: 1.1782942976878985e-07, 1.4848321370664053e-05, 0.0079355388879776
INFO - load Hessian from /workspace/hessians/quip-format/hessians/45_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:49:14 45.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:45.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 76 iterations, error 19831.3203125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 45.self_attn.o_proj 1st kmeans time: 29.021403312683105
INFO - 45.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 45.self_attn.o_proj proxy error before VPTQ: 2.7900080112885917e-06, 0.00014303166244644672, 0.01950622722506523
INFO - 45.self_attn.o_proj 1st error time: 2.2786245346069336
INFO - 45.self_attn.o_proj proxy error after VPTQ: 1.8662055936147226e-06, 0.00014303166244644672, 0.01304749958217144
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7514.392578125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 45.self_attn.o_proj residual time: 0.30937719345092773
INFO - 45.self_attn.o_proj 2ed gptq time: 2.5691375732421875
INFO - 45.self_attn.o_proj 2ed error time: 0.0010037422180175781
INFO - 45.self_attn.o_proj proxy error after residual VPTQ: 5.934249429628835e-07, 0.00014303166244644672, 0.004148906096816063
INFO - load Hessian from /workspace/hessians/quip-format/hessians/45_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:49:50 45.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:45.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 67248.515625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 45.mlp.gate_proj 1st kmeans time: 132.40166974067688
INFO - 45.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 45.mlp.gate_proj proxy error before VPTQ: 9.594576113158837e-06, 0.00023313771816901863, 0.04115411266684532
INFO - 45.mlp.gate_proj 1st error time: 4.728947401046753
INFO - 45.mlp.gate_proj proxy error after VPTQ: 7.217691745609045e-06, 0.00023313771816901863, 0.03095891885459423
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 94 iterations, error 23843.79296875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 45.mlp.gate_proj residual time: 0.9869251251220703
INFO - 45.mlp.gate_proj 2ed gptq time: 5.014002561569214
INFO - 45.mlp.gate_proj 2ed error time: 0.0011360645294189453
INFO - 45.mlp.gate_proj proxy error after residual VPTQ: 2.3462848730559926e-06, 0.00023313771816901863, 0.010063943453133106
INFO - load Hessian from /workspace/hessians/quip-format/hessians/45_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:52:16 45.mlp.up_proj
INFO - enabling norm dim 0, layer_name:45.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 72506.390625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 45.mlp.up_proj 1st kmeans time: 132.60416841506958
INFO - 45.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 45.mlp.up_proj proxy error before VPTQ: 8.388016794924624e-06, 0.00014179322170093656, 0.05915668234229088
INFO - 45.mlp.up_proj 1st error time: 4.713203191757202
INFO - 45.mlp.up_proj proxy error after VPTQ: 6.215633220563177e-06, 0.00014179322170093656, 0.043835896998643875
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 25584.79296875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 45.mlp.up_proj residual time: 1.0766196250915527
INFO - 45.mlp.up_proj 2ed gptq time: 5.0917088985443115
INFO - 45.mlp.up_proj 2ed error time: 0.0011134147644042969
INFO - 45.mlp.up_proj proxy error after residual VPTQ: 1.9990475266240537e-06, 0.00014179322170093656, 0.014098329469561577
INFO - load Hessian from /workspace/hessians/quip-format/hessians/45_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:54:44 45.mlp.down_proj
INFO - enabling norm dim 0, layer_name:45.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 82694.828125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 45.mlp.down_proj 1st kmeans time: 132.3878583908081
INFO - 45.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 45.mlp.down_proj proxy error before VPTQ: 6.202678832778474e-07, 1.1394852663215715e-05, 0.05443404242396355
INFO - 45.mlp.down_proj 1st error time: 8.73420524597168
INFO - 45.mlp.down_proj proxy error after VPTQ: 5.040993187321874e-07, 1.1394852663215715e-05, 0.044239211827516556
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 28865.19140625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 45.mlp.down_proj residual time: 1.0134310722351074
INFO - 45.mlp.down_proj 2ed gptq time: 9.513705492019653
INFO - 45.mlp.down_proj 2ed error time: 0.002424001693725586
INFO - 45.mlp.down_proj proxy error after residual VPTQ: 1.6582582418323e-07, 1.1394852663215715e-05, 0.01455269567668438
INFO - ----Quantizing layer 46 ...---- 2026-02-23 14:57:20 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/46_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:57:21 46.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:46.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 72 iterations, error 13292.556640625
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 46.self_attn.q_proj 1st kmeans time: 27.541935682296753
INFO - 46.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 46.self_attn.q_proj proxy error before VPTQ: 1.4579570688511012e-06, 4.8192239773925394e-05, 0.030252942815423012
INFO - 46.self_attn.q_proj 1st error time: 2.269892930984497
INFO - 46.self_attn.q_proj proxy error after VPTQ: 1.0175635907216929e-06, 4.8192239773925394e-05, 0.02111467719078064
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 5220.6962890625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 46.self_attn.q_proj residual time: 0.3192269802093506
INFO - 46.self_attn.q_proj 2ed gptq time: 2.5564119815826416
INFO - 46.self_attn.q_proj 2ed error time: 0.00025153160095214844
INFO - 46.self_attn.q_proj proxy error after residual VPTQ: 3.48295429830614e-07, 4.8192239773925394e-05, 0.007227209862321615
INFO - load Hessian from /workspace/hessians/quip-format/hessians/46_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:57:55 46.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:46.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 23 iterations, error 3131.93408203125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 46.self_attn.k_proj 1st kmeans time: 2.3592193126678467
INFO - 46.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 46.self_attn.k_proj proxy error before VPTQ: 2.6387607476863195e-07, 1.2525791134976316e-05, 0.02106661908328533
INFO - 46.self_attn.k_proj 1st error time: 1.5185298919677734
INFO - 46.self_attn.k_proj proxy error after VPTQ: 2.0584784010679869e-07, 1.2525791134976316e-05, 0.016433918848633766
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1277.109130859375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 46.self_attn.k_proj residual time: 0.0956575870513916
INFO - 46.self_attn.k_proj 2ed gptq time: 1.9633381366729736
INFO - 46.self_attn.k_proj 2ed error time: 0.00033783912658691406
INFO - 46.self_attn.k_proj proxy error after residual VPTQ: 6.901065319198096e-08, 1.2525791134976316e-05, 0.005509484559297562
INFO - load Hessian from /workspace/hessians/quip-format/hessians/46_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:58:03 46.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:46.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 3362.4296875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 46.self_attn.v_proj 1st kmeans time: 2.2783751487731934
INFO - 46.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 46.self_attn.v_proj proxy error before VPTQ: 5.368829079088755e-07, 1.4812294466537423e-05, 0.036245763301849365
INFO - 46.self_attn.v_proj 1st error time: 1.5034410953521729
INFO - 46.self_attn.v_proj proxy error after VPTQ: 4.2425395463396853e-07, 1.4812294466537423e-05, 0.028642015531659126
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1367.0283203125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 46.self_attn.v_proj residual time: 0.09292817115783691
INFO - 46.self_attn.v_proj 2ed gptq time: 1.965865135192871
INFO - 46.self_attn.v_proj 2ed error time: 0.0003612041473388672
INFO - 46.self_attn.v_proj proxy error after residual VPTQ: 1.395205231347063e-07, 1.4812294466537423e-05, 0.009419238194823265
INFO - load Hessian from /workspace/hessians/quip-format/hessians/46_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:58:10 46.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:46.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 73 iterations, error 21253.623046875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 46.self_attn.o_proj 1st kmeans time: 28.38087821006775
INFO - 46.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 46.self_attn.o_proj proxy error before VPTQ: 2.9387711037998088e-06, 0.0002301185013493523, 0.01277068629860878
INFO - 46.self_attn.o_proj 1st error time: 2.2738075256347656
INFO - 46.self_attn.o_proj proxy error after VPTQ: 1.9114595488645136e-06, 0.0002301185013493523, 0.008306413888931274
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7928.91064453125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 46.self_attn.o_proj residual time: 0.3080463409423828
INFO - 46.self_attn.o_proj 2ed gptq time: 2.557969570159912
INFO - 46.self_attn.o_proj 2ed error time: 0.0008969306945800781
INFO - 46.self_attn.o_proj proxy error after residual VPTQ: 6.052716798876645e-07, 0.0002301185013493523, 0.0026302607730031013
INFO - load Hessian from /workspace/hessians/quip-format/hessians/46_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 14:58:45 46.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:46.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 65888.0703125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 46.mlp.gate_proj 1st kmeans time: 132.9168803691864
INFO - 46.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 46.mlp.gate_proj proxy error before VPTQ: 9.351422704639845e-06, 0.00021487445337697864, 0.043520402163267136
INFO - 46.mlp.gate_proj 1st error time: 4.72434663772583
INFO - 46.mlp.gate_proj proxy error after VPTQ: 6.805928478570422e-06, 0.00021487445337697864, 0.03167397528886795
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 23444.0703125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 46.mlp.gate_proj residual time: 1.0863125324249268
INFO - 46.mlp.gate_proj 2ed gptq time: 5.01531720161438
INFO - 46.mlp.gate_proj 2ed error time: 0.0011444091796875
INFO - 46.mlp.gate_proj proxy error after residual VPTQ: 2.2096473912824877e-06, 0.00021487445337697864, 0.010283434763550758
INFO - load Hessian from /workspace/hessians/quip-format/hessians/46_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:01:11 46.mlp.up_proj
INFO - enabling norm dim 0, layer_name:46.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 70920.515625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 46.mlp.up_proj 1st kmeans time: 132.4078221321106
INFO - 46.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 46.mlp.up_proj proxy error before VPTQ: 8.781415090197697e-06, 0.00021077165729366243, 0.041663169860839844
INFO - 46.mlp.up_proj 1st error time: 4.769078969955444
INFO - 46.mlp.up_proj proxy error after VPTQ: 6.294289960351307e-06, 0.00021077165729366243, 0.02986307628452778
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 91 iterations, error 25110.06640625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 46.mlp.up_proj residual time: 0.9318671226501465
INFO - 46.mlp.up_proj 2ed gptq time: 5.054084300994873
INFO - 46.mlp.up_proj 2ed error time: 0.0011277198791503906
INFO - 46.mlp.up_proj proxy error after residual VPTQ: 2.0224324543960392e-06, 0.00021077165729366243, 0.009595371782779694
INFO - load Hessian from /workspace/hessians/quip-format/hessians/46_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:03:38 46.mlp.down_proj
INFO - enabling norm dim 0, layer_name:46.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 80324.421875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 46.mlp.down_proj 1st kmeans time: 132.44734954833984
INFO - 46.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 46.mlp.down_proj proxy error before VPTQ: 6.400024972208485e-07, 1.998869629460387e-05, 0.03201822191476822
INFO - 46.mlp.down_proj 1st error time: 9.160963773727417
INFO - 46.mlp.down_proj proxy error after VPTQ: 4.777105004905025e-07, 1.998869629460387e-05, 0.023899031803011894
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 28426.63671875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 46.mlp.down_proj residual time: 1.102855920791626
INFO - 46.mlp.down_proj 2ed gptq time: 9.619399547576904
INFO - 46.mlp.down_proj 2ed error time: 0.0026710033416748047
INFO - 46.mlp.down_proj proxy error after residual VPTQ: 1.5687642473949381e-07, 1.998869629460387e-05, 0.007848257198929787
INFO - ----Quantizing layer 47 ...---- 2026-02-23 15:06:15 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/47_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:06:16 47.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:47.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 73 iterations, error 15100.1796875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 47.self_attn.q_proj 1st kmeans time: 27.914854049682617
INFO - 47.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 47.self_attn.q_proj proxy error before VPTQ: 1.5501249208682566e-06, 5.642310861730948e-05, 0.02747322805225849
INFO - 47.self_attn.q_proj 1st error time: 2.3089191913604736
INFO - 47.self_attn.q_proj proxy error after VPTQ: 1.030289126902062e-06, 5.642310861730948e-05, 0.018260056152939796
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 5895.84130859375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 47.self_attn.q_proj residual time: 0.42487025260925293
INFO - 47.self_attn.q_proj 2ed gptq time: 2.555572509765625
INFO - 47.self_attn.q_proj 2ed error time: 0.00029397010803222656
INFO - 47.self_attn.q_proj proxy error after residual VPTQ: 3.4812617855095596e-07, 5.642310861730948e-05, 0.006169922184199095
INFO - load Hessian from /workspace/hessians/quip-format/hessians/47_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:06:51 47.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:47.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 22 iterations, error 3500.51953125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 47.self_attn.k_proj 1st kmeans time: 2.2758917808532715
INFO - 47.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 47.self_attn.k_proj proxy error before VPTQ: 2.8055379175384587e-07, 1.3549216419050936e-05, 0.02070627361536026
INFO - 47.self_attn.k_proj 1st error time: 1.541764259338379
INFO - 47.self_attn.k_proj proxy error after VPTQ: 2.083081511727869e-07, 1.3549216419050936e-05, 0.01537418458610773
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1450.3045654296875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 47.self_attn.k_proj residual time: 0.09907793998718262
INFO - 47.self_attn.k_proj 2ed gptq time: 1.9736042022705078
INFO - 47.self_attn.k_proj 2ed error time: 0.00047087669372558594
INFO - 47.self_attn.k_proj proxy error after residual VPTQ: 6.883620073949714e-08, 1.3549216419050936e-05, 0.005080456379801035
INFO - load Hessian from /workspace/hessians/quip-format/hessians/47_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:06:58 47.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:47.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 20 iterations, error 3715.1123046875
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 47.self_attn.v_proj 1st kmeans time: 2.094529151916504
INFO - 47.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 47.self_attn.v_proj proxy error before VPTQ: 7.701727327003027e-07, 2.4883516744012013e-05, 0.0309511199593544
INFO - 47.self_attn.v_proj 1st error time: 1.5097291469573975
INFO - 47.self_attn.v_proj proxy error after VPTQ: 5.913233280807617e-07, 2.4883516744012013e-05, 0.023763656616210938
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1533.924560546875
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 47.self_attn.v_proj residual time: 0.0934286117553711
INFO - 47.self_attn.v_proj 2ed gptq time: 1.9583778381347656
INFO - 47.self_attn.v_proj 2ed error time: 0.0003914833068847656
INFO - 47.self_attn.v_proj proxy error after residual VPTQ: 1.9486932956169767e-07, 2.4883516744012013e-05, 0.007831261493265629
INFO - load Hessian from /workspace/hessians/quip-format/hessians/47_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:07:05 47.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:47.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 70 iterations, error 19281.083984375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 47.self_attn.o_proj 1st kmeans time: 26.89350438117981
INFO - 47.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 47.self_attn.o_proj proxy error before VPTQ: 3.340390094308532e-06, 0.00019473840075079352, 0.017153216525912285
INFO - 47.self_attn.o_proj 1st error time: 2.3460309505462646
INFO - 47.self_attn.o_proj proxy error after VPTQ: 2.1790833670820575e-06, 0.00019473840075079352, 0.011189797893166542
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 7437.99365234375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 47.self_attn.o_proj residual time: 0.31522655487060547
INFO - 47.self_attn.o_proj 2ed gptq time: 2.6179375648498535
INFO - 47.self_attn.o_proj 2ed error time: 0.0012516975402832031
INFO - 47.self_attn.o_proj proxy error after residual VPTQ: 7.010193030509981e-07, 0.00019473840075079352, 0.0035997999366372824
INFO - load Hessian from /workspace/hessians/quip-format/hessians/47_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:07:39 47.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:47.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 65134.234375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 47.mlp.gate_proj 1st kmeans time: 132.40471386909485
INFO - 47.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 47.mlp.gate_proj proxy error before VPTQ: 9.719453373691067e-06, 0.0003338709066156298, 0.029111411422491074
INFO - 47.mlp.gate_proj 1st error time: 4.841230154037476
INFO - 47.mlp.gate_proj proxy error after VPTQ: 6.557243068527896e-06, 0.0003338709066156298, 0.019640054553747177
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 95 iterations, error 23310.470703125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 47.mlp.gate_proj residual time: 1.0654590129852295
INFO - 47.mlp.gate_proj 2ed gptq time: 5.028890371322632
INFO - 47.mlp.gate_proj 2ed error time: 0.0013484954833984375
INFO - 47.mlp.gate_proj proxy error after residual VPTQ: 2.1163236851862166e-06, 0.0003338709066156298, 0.00633874861523509
INFO - load Hessian from /workspace/hessians/quip-format/hessians/47_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:10:05 47.mlp.up_proj
INFO - enabling norm dim 0, layer_name:47.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 66292.96875
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 47.mlp.up_proj 1st kmeans time: 132.39953112602234
INFO - 47.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 47.mlp.up_proj proxy error before VPTQ: 1.2227121260366403e-05, 0.0004978757351636887, 0.024558579549193382
INFO - 47.mlp.up_proj 1st error time: 4.757737636566162
INFO - 47.mlp.up_proj proxy error after VPTQ: 8.636843631393276e-06, 0.0004978757351636887, 0.017347387969493866
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 93 iterations, error 23949.03515625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 47.mlp.up_proj residual time: 0.9638733863830566
INFO - 47.mlp.up_proj 2ed gptq time: 5.042744874954224
INFO - 47.mlp.up_proj 2ed error time: 0.0014123916625976562
INFO - 47.mlp.up_proj proxy error after residual VPTQ: 2.8048380045220256e-06, 0.0004978757351636887, 0.005633610766381025
INFO - load Hessian from /workspace/hessians/quip-format/hessians/47_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:12:32 47.mlp.down_proj
INFO - enabling norm dim 0, layer_name:47.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 79685.1640625
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 47.mlp.down_proj 1st kmeans time: 132.34970545768738
INFO - 47.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 47.mlp.down_proj proxy error before VPTQ: 6.340322897813166e-07, 1.624638753128238e-05, 0.03902604803442955
INFO - 47.mlp.down_proj 1st error time: 8.455281257629395
INFO - 47.mlp.down_proj proxy error after VPTQ: 4.4778832375413913e-07, 1.624638753128238e-05, 0.027562331408262253
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 28331.904296875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 47.mlp.down_proj residual time: 1.0101466178894043
INFO - 47.mlp.down_proj 2ed gptq time: 9.510398387908936
INFO - 47.mlp.down_proj 2ed error time: 0.0023522377014160156
INFO - 47.mlp.down_proj proxy error after residual VPTQ: 1.4722489538598893e-07, 1.624638753128238e-05, 0.009062008000910282
INFO - ----Quantizing layer 48 ...---- 2026-02-23 15:15:08 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/48_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:15:08 48.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:48.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 13833.484375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 48.self_attn.q_proj 1st kmeans time: 38.25291848182678
INFO - 48.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 48.self_attn.q_proj proxy error before VPTQ: 1.5013204119895818e-06, 5.4421972890850157e-05, 0.02758665941655636
INFO - 48.self_attn.q_proj 1st error time: 2.2867636680603027
INFO - 48.self_attn.q_proj proxy error after VPTQ: 9.288224873671425e-07, 5.4421972890850157e-05, 0.01706704869866371
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 5489.775390625
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 48.self_attn.q_proj residual time: 0.31043171882629395
INFO - 48.self_attn.q_proj 2ed gptq time: 2.629141092300415
INFO - 48.self_attn.q_proj 2ed error time: 0.0003666877746582031
INFO - 48.self_attn.q_proj proxy error after residual VPTQ: 3.184690058333217e-07, 5.4421972890850157e-05, 0.0058518461883068085
INFO - load Hessian from /workspace/hessians/quip-format/hessians/48_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:15:54 48.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:48.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 27 iterations, error 3290.44189453125
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 48.self_attn.k_proj 1st kmeans time: 2.745851516723633
INFO - 48.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 48.self_attn.k_proj proxy error before VPTQ: 2.3726983044980443e-07, 1.3828037481289357e-05, 0.01715860515832901
INFO - 48.self_attn.k_proj 1st error time: 1.5317327976226807
INFO - 48.self_attn.k_proj proxy error after VPTQ: 1.6629829246994632e-07, 1.3828037481289357e-05, 0.0120261674746871
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1368.1260986328125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 48.self_attn.k_proj residual time: 0.0911567211151123
INFO - 48.self_attn.k_proj 2ed gptq time: 1.9901340007781982
INFO - 48.self_attn.k_proj 2ed error time: 0.0003898143768310547
INFO - 48.self_attn.k_proj proxy error after residual VPTQ: 5.567535410477831e-08, 1.3828037481289357e-05, 0.004026265814900398
INFO - load Hessian from /workspace/hessians/quip-format/hessians/48_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:16:01 48.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:48.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 29 iterations, error 3553.81396484375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 48.self_attn.v_proj 1st kmeans time: 2.9401261806488037
INFO - 48.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 48.self_attn.v_proj proxy error before VPTQ: 7.842005516067729e-07, 3.1155519536696374e-05, 0.025170518085360527
INFO - 48.self_attn.v_proj 1st error time: 1.5104596614837646
INFO - 48.self_attn.v_proj proxy error after VPTQ: 5.716353825846454e-07, 3.1155519536696374e-05, 0.018347805365920067
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1460.7935791015625
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 48.self_attn.v_proj residual time: 0.09177088737487793
INFO - 48.self_attn.v_proj 2ed gptq time: 2.029165267944336
INFO - 48.self_attn.v_proj 2ed error time: 0.00042891502380371094
INFO - 48.self_attn.v_proj proxy error after residual VPTQ: 1.8812147573044058e-07, 3.1155519536696374e-05, 0.006038142833858728
INFO - load Hessian from /workspace/hessians/quip-format/hessians/48_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:16:10 48.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:48.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 79 iterations, error 20870.40234375
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 48.self_attn.o_proj 1st kmeans time: 30.349635124206543
INFO - 48.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 48.self_attn.o_proj proxy error before VPTQ: 3.7259430882841116e-06, 0.0007116373162716627, 0.00523573299869895
INFO - 48.self_attn.o_proj 1st error time: 2.370763063430786
INFO - 48.self_attn.o_proj proxy error after VPTQ: 2.2617768991040066e-06, 0.0007116373162716627, 0.0031782719306647778
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 8020.28857421875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 48.self_attn.o_proj residual time: 0.30962109565734863
INFO - 48.self_attn.o_proj 2ed gptq time: 2.5686352252960205
INFO - 48.self_attn.o_proj 2ed error time: 0.0009129047393798828
INFO - 48.self_attn.o_proj proxy error after residual VPTQ: 7.125063348212279e-07, 0.0007116373162716627, 0.001001221127808094
INFO - load Hessian from /workspace/hessians/quip-format/hessians/48_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:16:47 48.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:48.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 49556.74609375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 48.mlp.gate_proj 1st kmeans time: 133.49707698822021
INFO - 48.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 48.mlp.gate_proj proxy error before VPTQ: 1.6101639630505815e-05, 0.00865143071860075, 0.0018611534032970667
INFO - 48.mlp.gate_proj 1st error time: 4.730704307556152
INFO - 48.mlp.gate_proj proxy error after VPTQ: 1.0458880751684774e-05, 0.00865143071860075, 0.0012089192168787122
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 90 iterations, error 18274.29296875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 48.mlp.gate_proj residual time: 0.984816312789917
INFO - 48.mlp.gate_proj 2ed gptq time: 5.266702890396118
INFO - 48.mlp.gate_proj 2ed error time: 0.0012624263763427734
INFO - 48.mlp.gate_proj proxy error after residual VPTQ: 3.48799949279055e-06, 0.00865143071860075, 0.0004031702410429716
INFO - load Hessian from /workspace/hessians/quip-format/hessians/48_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:19:14 48.mlp.up_proj
INFO - enabling norm dim 0, layer_name:48.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 54066.3984375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 48.mlp.up_proj 1st kmeans time: 133.42613196372986
INFO - 48.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 48.mlp.up_proj proxy error before VPTQ: 1.529916698927991e-05, 0.001119957771152258, 0.013660485856235027
INFO - 48.mlp.up_proj 1st error time: 4.8138134479522705
INFO - 48.mlp.up_proj proxy error after VPTQ: 9.602645150152966e-06, 0.001119957771152258, 0.008574113249778748
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 81 iterations, error 19903.140625
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 48.mlp.up_proj residual time: 0.948401927947998
INFO - 48.mlp.up_proj 2ed gptq time: 5.130056619644165
INFO - 48.mlp.up_proj 2ed error time: 0.0012607574462890625
INFO - 48.mlp.up_proj proxy error after residual VPTQ: 3.1848662729316857e-06, 0.001119957771152258, 0.0028437378350645304
INFO - load Hessian from /workspace/hessians/quip-format/hessians/48_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:21:42 48.mlp.down_proj
INFO - enabling norm dim 0, layer_name:48.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 65247.51953125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 48.mlp.down_proj 1st kmeans time: 132.36717438697815
INFO - 48.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 48.mlp.down_proj proxy error before VPTQ: 6.800217420277477e-07, 3.28456808347255e-05, 0.02070353552699089
INFO - 48.mlp.down_proj 1st error time: 8.435195446014404
INFO - 48.mlp.down_proj proxy error after VPTQ: 4.782719429385907e-07, 3.28456808347255e-05, 0.014561181887984276
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 23732.58984375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 48.mlp.down_proj residual time: 1.0166516304016113
INFO - 48.mlp.down_proj 2ed gptq time: 9.516425848007202
INFO - 48.mlp.down_proj 2ed error time: 0.002536773681640625
INFO - 48.mlp.down_proj proxy error after residual VPTQ: 1.597671825948055e-07, 3.28456808347255e-05, 0.004864176269620657
INFO - ----Quantizing layer 49 ...---- 2026-02-23 15:24:18 on cuda:0 dtype torch.bfloat16
INFO - dict_keys(['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj', 'mlp.gate_proj', 'mlp.up_proj', 'mlp.down_proj'])
INFO - load Hessian from /workspace/hessians/quip-format/hessians/49_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:24:19 49.self_attn.q_proj
INFO - enabling norm dim 0, layer_name:49.self_attn.q_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 75 iterations, error 13803.779296875
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 49.self_attn.q_proj 1st kmeans time: 28.63801121711731
INFO - 49.self_attn.q_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 49.self_attn.q_proj proxy error before VPTQ: 1.4395266134670237e-06, 5.3827683586860076e-05, 0.026743238791823387
INFO - 49.self_attn.q_proj 1st error time: 2.3763532638549805
INFO - 49.self_attn.q_proj proxy error after VPTQ: 8.139581382238248e-07, 5.3827683586860076e-05, 0.015121552161872387
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 94 iterations, error 5390.7607421875
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 49.self_attn.q_proj residual time: 0.3928101062774658
INFO - 49.self_attn.q_proj 2ed gptq time: 2.677305221557617
INFO - 49.self_attn.q_proj 2ed error time: 0.00042128562927246094
INFO - 49.self_attn.q_proj proxy error after residual VPTQ: 2.765523845482676e-07, 5.3827683586860076e-05, 0.005137735046446323
INFO - load Hessian from /workspace/hessians/quip-format/hessians/49_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:24:55 49.self_attn.k_proj
INFO - enabling norm dim 0, layer_name:49.self_attn.k_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 25 iterations, error 3147.1572265625
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 49.self_attn.k_proj 1st kmeans time: 2.5521769523620605
INFO - 49.self_attn.k_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 49.self_attn.k_proj proxy error before VPTQ: 2.387946551607456e-07, 1.3260771083878353e-05, 0.01800759881734848
INFO - 49.self_attn.k_proj 1st error time: 1.5002977848052979
INFO - 49.self_attn.k_proj proxy error after VPTQ: 1.595496712525346e-07, 1.3260771083878353e-05, 0.012031704187393188
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1294.9923095703125
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 49.self_attn.k_proj residual time: 0.09029483795166016
INFO - 49.self_attn.k_proj 2ed gptq time: 1.9531633853912354
INFO - 49.self_attn.k_proj 2ed error time: 0.0004303455352783203
INFO - 49.self_attn.k_proj proxy error after residual VPTQ: 5.344134734741601e-08, 1.3260771083878353e-05, 0.004030033014714718
INFO - load Hessian from /workspace/hessians/quip-format/hessians/49_qkv.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:25:02 49.self_attn.v_proj
INFO - enabling norm dim 0, layer_name:49.self_attn.v_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([1024, 4096]), weights shape: torch.Size([1024, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 23 iterations, error 3309.25537109375
INFO - idx: 1, quant_data shape: torch.Size([524288, 8])
INFO - idx: 1, quant_data shape: torch.Size([1024, 4096])
INFO - quantized_data shape: torch.Size([1024, 4096])
INFO - 49.self_attn.v_proj 1st kmeans time: 2.3750197887420654
INFO - 49.self_attn.v_proj qweight init shape: torch.Size([1024, 4096]), weight shape: torch.Size([1024, 4096])
INFO - 49.self_attn.v_proj proxy error before VPTQ: 6.896403874634416e-07, 1.8107537471223623e-05, 0.03808581829071045
INFO - 49.self_attn.v_proj 1st error time: 1.552565336227417
INFO - 49.self_attn.v_proj proxy error after VPTQ: 4.873090801993385e-07, 1.8107537471223623e-05, 0.026911946013569832
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 1354.3275146484375
INFO - idx: 1, res quant_data shape: torch.Size([1024, 4096])
INFO - 49.self_attn.v_proj residual time: 0.09650468826293945
INFO - 49.self_attn.v_proj 2ed gptq time: 2.0496366024017334
INFO - 49.self_attn.v_proj 2ed error time: 0.0004956722259521484
INFO - 49.self_attn.v_proj proxy error after residual VPTQ: 1.6038680428209773e-07, 1.8107537471223623e-05, 0.00885746069252491
INFO - load Hessian from /workspace/hessians/quip-format/hessians/49_o.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:25:10 49.self_attn.o_proj
INFO - enabling norm dim 0, layer_name:49.self_attn.o_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 4096]), weights shape: torch.Size([4096, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 85 iterations, error 13996.251953125
INFO - idx: 1, quant_data shape: torch.Size([2097152, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 4096])
INFO - quantized_data shape: torch.Size([4096, 4096])
INFO - 49.self_attn.o_proj 1st kmeans time: 32.41622829437256
INFO - 49.self_attn.o_proj qweight init shape: torch.Size([4096, 4096]), weight shape: torch.Size([4096, 4096])
INFO - 49.self_attn.o_proj proxy error before VPTQ: 1.6662303096381947e-05, 0.02571174129843712, 0.0006480425945483148
INFO - 49.self_attn.o_proj 1st error time: 2.327794313430786
INFO - 49.self_attn.o_proj proxy error after VPTQ: 1.159829116659239e-05, 0.02571174129843712, 0.00045108929043635726
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 94 iterations, error 5724.30859375
INFO - idx: 1, res quant_data shape: torch.Size([4096, 4096])
INFO - 49.self_attn.o_proj residual time: 0.30318760871887207
INFO - 49.self_attn.o_proj 2ed gptq time: 2.687412977218628
INFO - 49.self_attn.o_proj 2ed error time: 0.0012784004211425781
INFO - 49.self_attn.o_proj proxy error after residual VPTQ: 4.032799552078359e-06, 0.02571174129843712, 0.00015684661047998816
INFO - load Hessian from /workspace/hessians/quip-format/hessians/49_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:25:49 49.mlp.gate_proj
INFO - enabling norm dim 0, layer_name:49.mlp.gate_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 61219.609375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 49.mlp.gate_proj 1st kmeans time: 132.39540648460388
INFO - 49.mlp.gate_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 49.mlp.gate_proj proxy error before VPTQ: 1.138409788836725e-05, 0.0008820031071081758, 0.01290709525346756
INFO - 49.mlp.gate_proj 1st error time: 4.830336570739746
INFO - 49.mlp.gate_proj proxy error after VPTQ: 5.709946435672464e-06, 0.0008820031071081758, 0.006473839282989502
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 86 iterations, error 21932.126953125
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 49.mlp.gate_proj residual time: 0.8850643634796143
INFO - 49.mlp.gate_proj 2ed gptq time: 5.033571720123291
INFO - 49.mlp.gate_proj 2ed error time: 0.0013735294342041016
INFO - 49.mlp.gate_proj proxy error after residual VPTQ: 1.8438208826410118e-06, 0.0008820031071081758, 0.002090492518618703
INFO - load Hessian from /workspace/hessians/quip-format/hessians/49_up.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:28:14 49.mlp.up_proj
INFO - enabling norm dim 0, layer_name:49.mlp.up_proj, scale:torch.Size([4096]), bias:torch.Size([4096])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([14336, 4096]), weights shape: torch.Size([14336, 4096])
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 101 iterations, error 60929.96484375
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([14336, 4096])
INFO - quantized_data shape: torch.Size([14336, 4096])
INFO - 49.mlp.up_proj 1st kmeans time: 132.56163024902344
INFO - 49.mlp.up_proj qweight init shape: torch.Size([14336, 4096]), weight shape: torch.Size([14336, 4096])
INFO - 49.mlp.up_proj proxy error before VPTQ: 1.4748350622539874e-05, 0.0027188132517039776, 0.005424554459750652
INFO - 49.mlp.up_proj 1st error time: 4.715227842330933
INFO - 49.mlp.up_proj proxy error after VPTQ: 6.3519282775814645e-06, 0.0027188132517039776, 0.00233628717251122
INFO - group_size: 4096 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 97 iterations, error 22167.85546875
INFO - idx: 1, res quant_data shape: torch.Size([14336, 4096])
INFO - 49.mlp.up_proj residual time: 0.9959402084350586
INFO - 49.mlp.up_proj 2ed gptq time: 5.042975664138794
INFO - 49.mlp.up_proj 2ed error time: 0.0012652873992919922
INFO - 49.mlp.up_proj proxy error after residual VPTQ: 2.0507936824287754e-06, 0.0027188132517039776, 0.0007542973617091775
INFO - load Hessian from /workspace/hessians/quip-format/hessians/49_down.pt
INFO - ----Quantizing llama ...---- 2026-02-23 15:30:42 49.mlp.down_proj
INFO - enabling norm dim 0, layer_name:49.mlp.down_proj, scale:torch.Size([14336]), bias:torch.Size([14336])
INFO - kmeans_mode: hessian, enable_perm: True, enable_norm: True, 
INFO - data shape: torch.Size([4096, 14336]), weights shape: torch.Size([4096, 14336])
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - cuml kmeans 95 iterations, error 44203.01953125
INFO - idx: 1, quant_data shape: torch.Size([7340032, 8])
INFO - idx: 1, quant_data shape: torch.Size([4096, 14336])
INFO - quantized_data shape: torch.Size([4096, 14336])
INFO - 49.mlp.down_proj 1st kmeans time: 125.94957542419434
INFO - 49.mlp.down_proj qweight init shape: torch.Size([4096, 14336]), weight shape: torch.Size([4096, 14336])
INFO - 49.mlp.down_proj proxy error before VPTQ: 5.514864938049868e-07, 4.471421198104508e-05, 0.012333584018051624
INFO - 49.mlp.down_proj 1st error time: 8.442019939422607
INFO - 49.mlp.down_proj proxy error after VPTQ: 3.2779666980786715e-07, 4.471421198104508e-05, 0.00733092799782753
INFO - group_size: 14336 number of groups: 1
INFO - idx: 0, num_centroids: -1, skip
INFO - kmeans_mode: hessian, cuml kmeans, 256 clusters
INFO - cuml kmeans 101 iterations, error 17013.564453125
INFO - idx: 1, res quant_data shape: torch.Size([4096, 14336])
INFO - 49.mlp.down_proj residual time: 1.0048174858093262
INFO - 49.mlp.down_proj 2ed gptq time: 9.514541625976562
INFO - 49.mlp.down_proj 2ed error time: 0.0026481151580810547
INFO - 49.mlp.down_proj proxy error after residual VPTQ: 1.1428689106196543e-07, 4.471421198104508e-05, 0.0025559409987181425
----quantization done ...---- 2026-02-23 15:33:11

Loading checkpoint shards:   0%|          | 0/5 [00:00<?, ?it/s]
Loading checkpoint shards:  20%|██        | 1/5 [00:05<00:22,  5.67s/it]
Loading checkpoint shards:  40%|████      | 2/5 [00:11<00:17,  5.73s/it]
Loading checkpoint shards:  60%|██████    | 3/5 [00:17<00:11,  5.72s/it]
Loading checkpoint shards:  80%|████████  | 4/5 [00:23<00:05,  5.87s/it]
Loading checkpoint shards: 100%|██████████| 5/5 [00:26<00:00,  4.91s/it]
Loading checkpoint shards: 100%|██████████| 5/5 [00:26<00:00,  5.29s/it]
qmodel: MistralForCausalLM(
  (model): MistralModel(
    (embed_tokens): Embedding(32128, 4096)
    (layers): ModuleList(
      (0-49): 50 x MistralDecoderLayer(
        (self_attn): MistralAttention(
          (q_proj): VQuantLinear(
            (centroids): Embedding(1, 524288)
            (res_centroids): Embedding(1, 2048)
          )
          (k_proj): VQuantLinear(
            (centroids): Embedding(1, 524288)
            (res_centroids): Embedding(1, 2048)
          )
          (v_proj): VQuantLinear(
            (centroids): Embedding(1, 524288)
            (res_centroids): Embedding(1, 2048)
          )
          (o_proj): VQuantLinear(
            (centroids): Embedding(1, 524288)
            (res_centroids): Embedding(1, 2048)
          )
        )
        (mlp): MistralMLP(
          (gate_proj): VQuantLinear(
            (centroids): Embedding(1, 524288)
            (res_centroids): Embedding(1, 2048)
          )
          (up_proj): VQuantLinear(
            (centroids): Embedding(1, 524288)
            (res_centroids): Embedding(1, 2048)
          )
          (down_proj): VQuantLinear(
            (centroids): Embedding(1, 524288)
            (res_centroids): Embedding(1, 2048)
          )
          (act_fn): SiLU()
        )
        (input_layernorm): MistralRMSNorm((4096,), eps=1e-05)
        (post_attention_layernorm): MistralRMSNorm((4096,), eps=1e-05)
      )
    )
    (norm): MistralRMSNorm((4096,), eps=1e-05)
    (rotary_emb): MistralRotaryEmbedding()
  )
  (lm_head): Linear(in_features=4096, out_features=32128, bias=False)
)
save config to /workspace/variant-e/output/2026-02-23-08-06-48/model/config.json
save model to /workspace/variant-e/output/2026-02-23-08-06-48/model/
save tokenizer to /workspace/variant-e/output/2026-02-23-08-06-48/model/
converting model indices from torch.uint16 to torch.uint16 as torch.int16
save packed model to /workspace/variant-e/output/2026-02-23-08-06-48/packed_model/
save tokenizer to /workspace/variant-e/output/2026-02-23-08-06-48/packed_model/

Replacing linear layers...:   0%|          | 0/657 [00:00<?, ?it/s]
Replacing linear layers...:  47%|████▋     | 311/657 [00:00<00:00, 3107.32it/s]
Replacing linear layers...:  95%|█████████▍| 623/657 [00:00<00:00, 3107.25it/s]
Replacing linear layers...: 100%|██████████| 657/657 [00:00<00:00, 1503.75it/s]
!!! Warning !!!: CUDA kernel not found, please check CUDA and VPTQ installation.
!!! Warning !!!: Running on Torch Implementation, which is extremely slow.
load packed model from /workspace/variant-e/output/2026-02-23-08-06-48/packed_model/
Loading dataset...

Generating test split:   0%|          | 0/4358 [00:00<?, ? examples/s]
Generating test split: 100%|██████████| 4358/4358 [00:00<00:00, 354507.80 examples/s]

Generating train split:   0%|          | 0/36718 [00:00<?, ? examples/s]
Generating train split: 100%|██████████| 36718/36718 [00:00<00:00, 856800.45 examples/s]

Generating validation split:   0%|          | 0/3760 [00:00<?, ? examples/s]
Generating validation split: 100%|██████████| 3760/3760 [00:00<00:00, 546981.93 examples/s]
Preprocessing dataset....
wikitext2
----Evaluating llama ...---- 2026-02-23 15:35:43

  0%|          | 0/50 [00:00<?, ?it/s]
  2%|▏         | 1/50 [00:03<02:41,  3.30s/it]
  4%|▍         | 2/50 [00:06<02:36,  3.26s/it]
  6%|▌         | 3/50 [00:09<02:32,  3.24s/it]
  8%|▊         | 4/50 [00:12<02:28,  3.24s/it]
 10%|█         | 5/50 [00:16<02:25,  3.23s/it]
 12%|█▏        | 6/50 [00:19<02:22,  3.23s/it]
 14%|█▍        | 7/50 [00:22<02:18,  3.23s/it]
 16%|█▌        | 8/50 [00:25<02:15,  3.23s/it]
 18%|█▊        | 9/50 [00:29<02:12,  3.23s/it]
 20%|██        | 10/50 [00:32<02:09,  3.23s/it]
 22%|██▏       | 11/50 [00:35<02:05,  3.23s/it]
 24%|██▍       | 12/50 [00:38<02:02,  3.23s/it]
 26%|██▌       | 13/50 [00:42<01:59,  3.23s/it]
 28%|██▊       | 14/50 [00:45<01:56,  3.23s/it]
 30%|███       | 15/50 [00:48<01:52,  3.23s/it]
 32%|███▏      | 16/50 [00:51<01:49,  3.23s/it]
 34%|███▍      | 17/50 [00:54<01:46,  3.23s/it]
 36%|███▌      | 18/50 [00:58<01:43,  3.22s/it]
 38%|███▊      | 19/50 [01:01<01:40,  3.23s/it]
 40%|████      | 20/50 [01:04<01:36,  3.23s/it]
 42%|████▏     | 21/50 [01:07<01:33,  3.23s/it]
 44%|████▍     | 22/50 [01:11<01:30,  3.23s/it]
 46%|████▌     | 23/50 [01:14<01:27,  3.23s/it]
 48%|████▊     | 24/50 [01:17<01:24,  3.23s/it]
 50%|█████     | 25/50 [01:20<01:20,  3.24s/it]
 52%|█████▏    | 26/50 [01:24<01:17,  3.24s/it]
 54%|█████▍    | 27/50 [01:27<01:14,  3.24s/it]
 56%|█████▌    | 28/50 [01:30<01:11,  3.26s/it]
 58%|█████▊    | 29/50 [01:33<01:08,  3.27s/it]
 60%|██████    | 30/50 [01:37<01:05,  3.27s/it]
 62%|██████▏   | 31/50 [01:40<01:01,  3.26s/it]
 64%|██████▍   | 32/50 [01:43<00:58,  3.26s/it]
 66%|██████▌   | 33/50 [01:46<00:55,  3.27s/it]
 68%|██████▊   | 34/50 [01:50<00:52,  3.28s/it]
 70%|███████   | 35/50 [01:53<00:49,  3.28s/it]
 72%|███████▏  | 36/50 [01:56<00:45,  3.27s/it]
 74%|███████▍  | 37/50 [02:00<00:42,  3.27s/it]
 76%|███████▌  | 38/50 [02:03<00:39,  3.27s/it]
 78%|███████▊  | 39/50 [02:06<00:35,  3.26s/it]
 80%|████████  | 40/50 [02:09<00:32,  3.26s/it]
 82%|████████▏ | 41/50 [02:13<00:29,  3.26s/it]
 84%|████████▍ | 42/50 [02:16<00:26,  3.25s/it]
 86%|████████▌ | 43/50 [02:19<00:22,  3.25s/it]
 88%|████████▊ | 44/50 [02:22<00:19,  3.25s/it]
 90%|█████████ | 45/50 [02:25<00:16,  3.24s/it]
 92%|█████████▏| 46/50 [02:29<00:12,  3.24s/it]
 94%|█████████▍| 47/50 [02:32<00:09,  3.24s/it]
 96%|█████████▌| 48/50 [02:35<00:06,  3.24s/it]
 98%|█████████▊| 49/50 [02:38<00:03,  3.24s/it]
100%|██████████| 50/50 [02:42<00:00,  3.24s/it]
100%|██████████| 50/50 [02:42<00:00,  3.24s/it]
7.81787633895874
Setting num_proc from 48 back to 1 for the train split to disable multiprocessing as it only contains one shard.
WARNING - Setting num_proc from 48 back to 1 for the train split to disable multiprocessing as it only contains one shard.

Generating train split: 0 examples [00:00, ? examples/s]
Generating train split: 4435 examples [00:00, 34776.88 examples/s]
Generating train split: 13602 examples [00:00, 57143.71 examples/s]
Generating train split: 22590 examples [00:00, 65252.96 examples/s]
Generating train split: 32053 examples [00:00, 70506.48 examples/s]
Generating train split: 41546 examples [00:00, 73770.40 examples/s]
Generating train split: 50510 examples [00:00, 74052.84 examples/s]
Generating train split: 59531 examples [00:00, 74834.27 examples/s]
Generating train split: 68750 examples [00:00, 76045.84 examples/s]
Generating train split: 77743 examples [00:01, 74572.08 examples/s]
Generating train split: 86966 examples [00:01, 74498.85 examples/s]
Generating train split: 95787 examples [00:01, 74859.93 examples/s]
Generating train split: 104859 examples [00:01, 74935.72 examples/s]
Generating train split: 113857 examples [00:01, 74975.87 examples/s]
Generating train split: 122789 examples [00:01, 76066.01 examples/s]
Generating train split: 132051 examples [00:01, 78113.18 examples/s]
Generating train split: 141177 examples [00:01, 79256.09 examples/s]
Generating train split: 150440 examples [00:02, 79924.54 examples/s]
Generating train split: 159772 examples [00:02, 81025.75 examples/s]
Generating train split: 168803 examples [00:02, 80056.01 examples/s]
Generating train split: 177746 examples [00:02, 79846.56 examples/s]
Generating train split: 186637 examples [00:02, 79364.53 examples/s]
Generating train split: 195882 examples [00:02, 80066.64 examples/s]
Generating train split: 204917 examples [00:02, 80100.66 examples/s]
Generating train split: 214014 examples [00:02, 80576.19 examples/s]
Generating train split: 223006 examples [00:02, 80058.14 examples/s]
Generating train split: 232246 examples [00:03, 80428.63 examples/s]
Generating train split: 241218 examples [00:03, 80217.67 examples/s]
Generating train split: 250396 examples [00:03, 80172.02 examples/s]
Generating train split: 259470 examples [00:03, 80255.18 examples/s]
Generating train split: 268486 examples [00:03, 79773.53 examples/s]
Generating train split: 277868 examples [00:03, 80635.08 examples/s]
Generating train split: 286955 examples [00:03, 79898.15 examples/s]
Generating train split: 295901 examples [00:03, 79730.43 examples/s]
Generating train split: 304853 examples [00:03, 79734.83 examples/s]
Generating train split: 314057 examples [00:04, 80452.97 examples/s]
Generating train split: 323150 examples [00:04, 80323.84 examples/s]
Generating train split: 332325 examples [00:04, 80176.10 examples/s]
Generating train split: 341256 examples [00:04, 79945.23 examples/s]
Generating train split: 350356 examples [00:04, 79943.01 examples/s]
Generating train split: 356317 examples [00:04, 77545.37 examples/s]
Setting num_proc from 48 back to 1 for the validation split to disable multiprocessing as it only contains one shard.
WARNING - Setting num_proc from 48 back to 1 for the validation split to disable multiprocessing as it only contains one shard.

Generating validation split: 0 examples [00:00, ? examples/s]
Generating validation split: 4640 examples [00:00, 44001.24 examples/s]
Generating validation split: 13718 examples [00:00, 59174.70 examples/s]
Generating validation split: 22910 examples [00:00, 68233.20 examples/s]
Generating validation split: 32282 examples [00:00, 72869.41 examples/s]
Generating validation split: 41615 examples [00:00, 74036.14 examples/s]
Generating validation split: 45576 examples [00:00, 70531.88 examples/s]
c4-new
----Evaluating llama ...---- 2026-02-23 15:38:56

  0%|          | 0/50 [00:00<?, ?it/s]
  2%|▏         | 1/50 [00:05<04:08,  5.07s/it]
  4%|▍         | 2/50 [00:10<04:04,  5.09s/it]
  6%|▌         | 3/50 [00:15<03:59,  5.09s/it]
  8%|▊         | 4/50 [00:20<03:54,  5.10s/it]
 10%|█         | 5/50 [00:25<03:49,  5.10s/it]
 12%|█▏        | 6/50 [00:30<03:44,  5.10s/it]
 14%|█▍        | 7/50 [00:35<03:39,  5.11s/it]
 16%|█▌        | 8/50 [00:40<03:34,  5.11s/it]
 18%|█▊        | 9/50 [00:45<03:29,  5.11s/it]
 20%|██        | 10/50 [00:51<03:24,  5.11s/it]
 22%|██▏       | 11/50 [00:56<03:19,  5.11s/it]
 24%|██▍       | 12/50 [01:01<03:14,  5.11s/it]
 26%|██▌       | 13/50 [01:06<03:09,  5.11s/it]
 28%|██▊       | 14/50 [01:11<03:03,  5.11s/it]
 30%|███       | 15/50 [01:16<02:58,  5.11s/it]
 32%|███▏      | 16/50 [01:21<02:53,  5.11s/it]
 34%|███▍      | 17/50 [01:26<02:48,  5.11s/it]
 36%|███▌      | 18/50 [01:31<02:43,  5.11s/it]
 38%|███▊      | 19/50 [01:37<02:38,  5.11s/it]
 40%|████      | 20/50 [01:42<02:33,  5.10s/it]
 42%|████▏     | 21/50 [01:47<02:27,  5.10s/it]
 44%|████▍     | 22/50 [01:52<02:23,  5.11s/it]
 46%|████▌     | 23/50 [01:57<02:18,  5.13s/it]
 48%|████▊     | 24/50 [02:02<02:13,  5.13s/it]
 50%|█████     | 25/50 [02:07<02:08,  5.14s/it]
 52%|█████▏    | 26/50 [02:12<02:03,  5.14s/it]
 54%|█████▍    | 27/50 [02:18<01:58,  5.14s/it]
 56%|█████▌    | 28/50 [02:23<01:52,  5.13s/it]
 58%|█████▊    | 29/50 [02:28<01:47,  5.12s/it]
 60%|██████    | 30/50 [02:33<01:42,  5.12s/it]
 62%|██████▏   | 31/50 [02:38<01:37,  5.11s/it]
 64%|██████▍   | 32/50 [02:43<01:31,  5.10s/it]
 66%|██████▌   | 33/50 [02:48<01:26,  5.10s/it]
 68%|██████▊   | 34/50 [02:53<01:21,  5.10s/it]
 70%|███████   | 35/50 [02:58<01:16,  5.10s/it]
 72%|███████▏  | 36/50 [03:03<01:11,  5.09s/it]
 74%|███████▍  | 37/50 [03:09<01:06,  5.10s/it]
 76%|███████▌  | 38/50 [03:14<01:01,  5.11s/it]
 78%|███████▊  | 39/50 [03:19<00:56,  5.10s/it]
 80%|████████  | 40/50 [03:24<00:51,  5.10s/it]
 82%|████████▏ | 41/50 [03:29<00:46,  5.11s/it]
 84%|████████▍ | 42/50 [03:34<00:40,  5.11s/it]
 86%|████████▌ | 43/50 [03:39<00:35,  5.12s/it]
 88%|████████▊ | 44/50 [03:44<00:30,  5.11s/it]
 90%|█████████ | 45/50 [03:49<00:25,  5.11s/it]
 92%|█████████▏| 46/50 [03:55<00:20,  5.12s/it]
 94%|█████████▍| 47/50 [04:00<00:15,  5.12s/it]
 96%|█████████▌| 48/50 [04:05<00:10,  5.12s/it]
 98%|█████████▊| 49/50 [04:10<00:05,  5.13s/it]
100%|██████████| 50/50 [04:15<00:00,  5.13s/it]
100%|██████████| 50/50 [04:15<00:00,  5.11s/it]
15.322999000549316