[CI] Test wrapq examples

[dayo09]

What?

We should provide wrapq example code's CI integration test.

[mhs4670go]

Directories

.
├── compare_ppl.py
├── debug_quant_outputs.py
├── __init__.py
├── llama
│   ├── __init__.py
│   ├── quantize_attn_decode.py
│   ├── quantize_attn_prefill.py
│   ├── quantize_decoder_layer_decode.py
│   ├── quantize_decoder_layer_prefill.py
│   └── quantize_mlp.py
├── nn
│   ├── quantize_conv3d.py
│   └── quantize_conv3d_special_case.py
├── quantize_full_qmodel_with_gptq.py
├── quantize_linear.py
├── quantize_with_gptq.py
└── qwen
    ├── __init__.py
    ├── quantize_qwen_text_mlp.py
    ├── quantize_text_attn.py
    ├── quantize_vision_attn.py
    ├── quantize_vision_block.py
    ├── quantize_vision_mlp.py
    ├── quantize_vision_patch_embed.py
    └── quantize_vision_patch_merger.py

Outputs

python tico/quantization/wrapq/examples/llama/quantize_decoder_layer_prefill.py
┌───────────── Quantization Error Summary ─────────────
│ Mean |diff|: 0.020452
│ PEIR       : 1.983250 %
└──────────────────────────────────────────────────────
     ┌───────────────────────────────────────────┐
 22.5┤                                           │
     │                                        •  │
     │                                           │
 16.7┤                                     •     │
     │                                           │
     │                                           │
     │                                           │
 10.9┤                                           │
     │                                           │
     │                        •                  │
  5.1┤                      ••                   │
     │                     •                     │
     │                                           │
     │               ••                          │
 -0.6┤              ••                           │
     │                                           │
     │         •                                 │
 -6.4┤        •                                  │
     │      •                                    │
     │    •                                      │
     │  •                                        │
-12.2┤                                           │
     └┬──────────┬─────────┬──────────┬─────────┬┘
    -12.2      -3.5       5.1       13.8     22.5 

Quantized Circle model saved to /home/seongwoo/TICO/decoder_layer_prefill.q.circle

[mhs4670go]

Before doing the work, let me show current status.

wrapq example scripts (PEIR)

Script	Model	PEIR (%)
quantize_decoder_layer_prefill.py	TinyLlama	46.538
quantize_decoder_layer_prefill.py	Llama-3.2-1B-Instruct	1.983
quantize_decoder_layer_decode.py	TinyLlama	12.592, 54.965, 51.108
quantize_decoder_layer_decode.py	Llama-3.2-1B-Instruct	2.737, 1.551, 2.007
quantize_attn_prefill.py	TinyLlama	19.569
quantize_attn_prefill.py	Llama-3.2-1B-Instruct	25.908
quantize_attn_decode.py	TinyLlama	3.966
quantize_attn_decode.py	Llama-3.2-1B-Instruct	9.357
quantize_mlp.py	TinyLlama	0.007
quantize_mlp.py	Llama-3.2-1B-Instruct	0.005
quantize_linear.py	nn	0.656
quantize_conv3d.py	nn	1.231
quantize_conv3d_special_case.py	nn	0.634
quantize_text_attn.py	Qwen3-VL-2B-Instruct	7.624
quantize_text_mlp.py	Qwen3-VL-2B-Instruct	2.837
quantize_vision_attn.py	Qwen3-VL-2B-Instruct	21.824
quantize_vision_block.py	Qwen3-VL-2B-Instruct	8.473
quantize_vision_mlp.py	Qwen3-VL-2B-Instruct	2.336
quantize_vision_patch_embed.py	Qwen3-VL-2B-Instruct	0.829
quantize_vision_patch_merger.py	Qwen3-VL-2B-Instruct	3.556

Notes

• PEIR = Peak Error Interval Ratio (lower is better)
  • max(|a - b|) / (max(a) - min(a))