[wrapq] High PEIR (46%) of LlamaDecoderLayer of TinyLlama

[dayo09]

What?

tico/quantization/wrapq/examples/llama/quantize_llama_decoder_layer.py PEIR is very high >=46%

Copied from #490 (comment)

@mhs4670go

I found that PEIR is very high. Is it okay?

ai-edge-torch                0.7.1
torch                        2.10.0+cpu
torch_xla2                   0.0.1.dev202412041639
torchaudio                   2.10.0+cpu
torchcodec                   0.10.0
torchvision                  0.25.0+cpu
transformers                 4.57.3


┌───────────── Quantization Error Summary ─────────────
│ Mean |diff|: 0.262138
│ PEIR       : 46.538607 %
└──────────────────────────────────────────────────────
     ┌───────────────────────────────────────────┐
 0.95┤                                           │
     │                                           │
     │                                           │
 0.60┤                                           │
     │                                           │
     │                                           │
     │                                           │
 0.25┤                                           │
     │            •  •  •••••••••••••••••••• ••  │
     │             •• •••••• ••••••••••••••      │
-0.09┤     • •  •••••••• •••••••••••••           │
     │  •              •                         │
     │                                           │
-0.44┤                                           │
     │                                           │
     │                                           │
     │                                           │
-0.79┤                                           │
     │                                           │
     │                                           │
-1.14┤                                           │
     └┬──────────┬─────────┬──────────┬─────────┬┘
    -1.14      -0.61     -0.09      0.43     0.95 

[mhs4670go]

First, the high PEIR doesn't come from the PR. The PEIR is low with Llama 3.2 instead of TinyLlama. So, let's investigate the causes with TinyLlama.