[quantization] Implement llama wrappers for decoding

[mhs4670go]

This commit implements llama wrappers for decoding.

TICO-DCO-1.0-Signed-off-by: seongwoo mhs4670go@naver.com

[mhs4670go]

┌──────────── Quantization Error Summary (Decode DecoderLayer / Random) ────────────
│ Mean |diff| (hidden): 0.037206
│ PEIR        (hidden): 3.549647 %
│ Mean |diff| (new_k) : 0.003758
│ PEIR        (new_k) : 1.324077 %
│ Mean |diff| (new_v) : 0.001980
│ PEIR        (new_v) : 1.693285 %
└──────────────────────────────────────────────────────────────────────────────────
    ┌────────────────────────────────────────────┐
 4.2┤                                            │
    │                                         •  │
    │                                      ••    │
 2.8┤                                    •••     │
    │                                  •••       │
    │                               ••••         │
    │                             ••••           │
 1.5┤                           ••••             │
    │                         ••••               │
    │                       ••••                 │
 0.2┤                     ••••                   │
    │                   ••••                     │
    │                 ••••                       │
    │               ••••                         │
-1.1┤             ••••                           │
    │           ••••                             │
    │         •••                                │
-2.5┤       ••••                                 │
    │     •••                                    │
    │   •••                                      │
    │  •                                         │
-3.8┤                                            │
    └┬──────────┬──────────┬─────────┬──────────┬┘
   -3.8       -1.8        0.2       2.2       4.2

[dayo09]

Suggested change

	variant = getattr(qcfg, "wrapper_variant", "prefill") if qcfg else "prefill"
	variant = getattr(qcfg, "wrapper_variant", "prefill") if qcfg else None

Quantized nn module may not be one of prefill/decode stage

[mhs4670go]

I desinged to use "prefill" as a default value. Do you think exact default value is needed?

Resolution Order
    ----------------
    1. Exact variant match
    2. "prefill" fallback
    3. Any registered variant (last-resort compatibility)

[mhs4670go]

Previously, lookup(fp_cls) was used without any notion of variants. Even after introducing variants, it is likely that most wrappers remain variant-agnostic, meaning only a single implementation exists as you said.

In this situation, requesting variant="decode" does not necessarily mean that every module must provide a dedicated decode-specific wrapper. Many modules (e.g., Linear, LayerNorm) have identical behavior between prefill and decode execution and therefore do not require separate implementations.

With fallback enabled, these variant-agnostic modules can be safely reused while still constructing a decode-specialized graph. This allows decode graph construction to proceed without forcing unnecessary wrapper duplication for modules whose behavior does not actually differ across variants.

[dayo09]

Thanks for the detailed explanation!

As you mentioned, most of them are irrelevant to the field 'variant'.
So I assumed that default value should be None for most cases and decode/prefill could be selected when they are relevant.

[dayo09]

Using a technical term, it's a leaky abstraction. Let's choose to be 'opt-in' variance. Let modules to have 'None' or 'invariant' if they don't belong to either of 'prefill' or 'decode.

[mhs4670go]

python tico/quantization/wrapq/examples/llama/quantize_attn_decode.py
┌───────────── Quantization Error Summary (Decode Attn / Random) ─────────────
│ Mean |diff|: 0.087903
│ PEIR       : 10.583212 %
└────────────────────────────────────────────────────────────────────────────
     ┌───────────────────────────────────────────┐
 2.13┤                                           │
     │                                        •  │
     │                                     •     │
 1.21┤                                  ••• •    │
     │                               ••••••      │
     │                             ••••••        │
     │                          ••••••••         │
 0.30┤                         •••••••           │
     │                       •••••••             │
     │                    • •••••••              │
-0.62┤                   •••••••                 │
     │                   •••••                   │
     │                 •••••                     │
     │                •••                        │
-1.53┤              •••                          │
     │                                           │
     │                                           │
-2.45┤                                           │
     │                                           │
     │  •                                        │
     │                                           │
-3.36┤                                           │
     └┬──────────┬─────────┬──────────┬─────────┬┘
    -3.4       -2.0      -0.6        0.8      2.1 

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

[mhs4670go]

@dayo09 I've introduced "common" variant. PTAL:)

[dayo09]

LGTM Thanks!