epic-quant / sweep_stdout.txt
toxzak's picture
Add 1.58/3/4/16-bit sweep, packed weights, real SDPA, COMPARISON.md
7e372c7
Raw
History Blame Contribute Delete
8.32 kB
��python : C:\Users\Zwmar\projects\e4b\epic_quant\loader.py:95: UserWarning: The given buffer is not writable, and
PyTorch does not support non-writable tensors. This means you can write to the underlying (supposedly non-writable)
buffer using the tensor. You may want to copy the buffer to protect its data or make it writable before converting it
to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at
C:\actions-runner\_work\pytorch\pytorch\pytorch\torch\csrc\utils\tensor_new.cpp:1592.)
At line:1 char:82
+ ... jects\e4b"; python -m epic_quant.bench --sweep --n-tokens 200 --out " ...
+ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+ CategoryInfo : NotSpecified: (C:\Users\Zwmar\..._new.cpp:1592.):String) [], RemoteException
+ FullyQualifiedErrorId : NativeCommandError
t = torch.frombuffer(row_view, dtype=torch_dtype)
[bench] locating model...
[bench] model at C:\Users\Zwmar\.lmstudio\hub\models--google--gemma-4-E4B\snapshots\a24c9379fd3839ae84e97f0b6aa3152fce9bd033\model.safetensors (15.99 GB)
[sweep] building FP16-baseline engine for PLE workload (warmup)...
[sweep] PLE hot hit rate: 84.5%, hot table MB: 107.5, lookups/sec: 21891
========== POLICY: 1.58bit (brief) (sliding=2, global=4, mlp=4) ==========
attn unquant: 1284.5 MB -> packed: 207.0 MB (saved 1077.6 MB)
mlp unquant: 6606.0 MB -> packed: 1653.4 MB (saved 4952.6 MB)
ple unquant: 110.1 MB -> packed: 27.8 MB (saved 82.3 MB)
sliding L2 recon: ple_gate=0.200, ple_proj=0.162, attn_q=1.114, attn_k=1.116, attn_v=1.116, attn_o=1.108, mlp_gate=0.186, mlp_up=0.174, mlp_down=0.204
global L2 recon: ple_gate=0.275, ple_proj=0.117, attn_q=0.195, attn_k=0.262, attn_v=0.173, attn_o=0.186, mlp_gate=0.229, mlp_up=0.220, mlp_down=0.214
sliding total ms: 2101, global total ms: 2128
sliding layer packed bytes: 43.3 MB
global layer packed bytes: 53.2 MB
========== POLICY: 3bit (sliding=3, global=4, mlp=4) ==========
attn unquant: 1284.5 MB -> packed: 321.6 MB (saved 962.9 MB)
mlp unquant: 6606.0 MB -> packed: 1653.4 MB (saved 4952.6 MB)
ple unquant: 110.1 MB -> packed: 27.8 MB (saved 82.3 MB)
sliding L2 recon: ple_gate=0.200, ple_proj=0.162, attn_q=0.302, attn_k=0.289, attn_v=0.285, attn_o=0.291, mlp_gate=0.186, mlp_up=0.174, mlp_down=0.204
global L2 recon: ple_gate=0.275, ple_proj=0.117, attn_q=0.195, attn_k=0.262, attn_v=0.173, attn_o=0.186, mlp_gate=0.229, mlp_up=0.220, mlp_down=0.214
sliding total ms: 2013, global total ms: 2130
sliding layer packed bytes: 46.6 MB
global layer packed bytes: 53.2 MB
========== POLICY: 4bit (uniform) (sliding=4, global=4, mlp=4) ==========
attn unquant: 1284.5 MB -> packed: 321.6 MB (saved 962.9 MB)
mlp unquant: 6606.0 MB -> packed: 1653.4 MB (saved 4952.6 MB)
ple unquant: 110.1 MB -> packed: 27.8 MB (saved 82.3 MB)
sliding L2 recon: ple_gate=0.200, ple_proj=0.162, attn_q=0.173, attn_k=0.165, attn_v=0.163, attn_o=0.166, mlp_gate=0.186, mlp_up=0.174, mlp_down=0.204
global L2 recon: ple_gate=0.275, ple_proj=0.117, attn_q=0.195, attn_k=0.262, attn_v=0.173, attn_o=0.186, mlp_gate=0.229, mlp_up=0.220, mlp_down=0.214
sliding total ms: 2106, global total ms: 2088
sliding layer packed bytes: 46.6 MB
global layer packed bytes: 53.2 MB
========== POLICY: 16bit (no quant) (sliding=16, global=16, mlp=16) ==========
attn unquant: 1284.5 MB -> packed: 1284.5 MB (saved 0.0 MB)
mlp unquant: 6606.0 MB -> packed: 6606.0 MB (saved 0.0 MB)
ple unquant: 110.1 MB -> packed: 110.1 MB (saved 0.0 MB)
sliding L2 recon: ple_gate=0.000, ple_proj=0.000, attn_q=0.000, attn_k=0.000, attn_v=0.000, attn_o=0.000, mlp_gate=0.000, mlp_up=0.000, mlp_down=0.000
global L2 recon: ple_gate=0.000, ple_proj=0.000, attn_q=0.000, attn_k=0.000, attn_v=0.000, attn_o=0.000, mlp_gate=0.000, mlp_up=0.000, mlp_down=0.000
sliding total ms: 839, global total ms: 740
sliding layer packed bytes: 186.1 MB
global layer packed bytes: 212.3 MB
[bench] wrote C:\Users\Zwmar\projects\e4b\sweep.json