[quantization] Fallback for complex models

[stamalakhov]

This PR:

1. quantizes the whole model (e.g. vlm) as a single layer in GPTQ
2. adds example script for general vlm models

log of `python tico/quantization/wrapq/examples/quantize_full_vlm_model_with_gptq.py --model HuggingFaceTB/SmolVLM-256M-Instruct --nsamples_for_qcalibration 20 --eval_tasks vqav2 --gptq_mse mse`

Namespace(model='HuggingFaceTB/SmolVLM-256M-Instruct', device='cuda', dtype='float32', seed=42, trust_remote_code=False, hf_token=None, cache_dir='/mnt/storage/transformers_cache', nsamples_for_qcalibration=20, nsamples_for_evaluation=50, calib_seq_len=2048, max_seq_len=2048, linear_weight_bits=4, gptq_mse='mse', eval_tasks='vqav2', sensitivity_path=None)
=== Config ===
Model            : HuggingFaceTB/SmolVLM-256M-Instruct
Device           : cuda
DType            : float32
Calib seq len    : 2048
Max seq len      : 2048

Loading FP model …
Evaluating original model
Evaluating original model
vqav2: EM=0.5600  (n=50)
Original EM: 0.5600  (dataset=vqav2, n=50)
Resolving data files: 100%|█████████████████████████████████████████████████████████████| 68/68 [00:00<00:00, 7662.06it/s]
Resolving data files: 100%|█████████████████████████████████████████████████████████████| 36/36 [00:00<00:00, 6735.73it/s]
Resolving data files: 100%|██████████████████████████████████████████████████████████| 143/143 [00:00<00:00, 10294.45it/s]
Resolving data files: 100%|█████████████████████████████████████████████████████████████| 68/68 [00:00<00:00, 6965.07it/s]
Resolving data files: 100%|█████████████████████████████████████████████████████████████| 36/36 [00:00<00:00, 7293.38it/s]
Resolving data files: 100%|██████████████████████████████████████████████████████████| 143/143 [00:00<00:00, 13307.27it/s]
[info] Loaded dataset: lmms-lab/VQAv2 (testdev, streaming), size=20
Applying GPTQ …
Quantizing layers: 100%|████████████████████████████████████████████████████████████████| 1/1 [02:02<00:00, 122.53s/layer]
Evaluating quantized model                                                                                                
vqav2: EM=0.5600  (n=50)
|model|vqav2|
|--|--|
|original|0.5600|
|quantized|0.5600|

Please see other results in #559
Draft: #559
Related: #548

TICO-DCO-1.0-Signed-off-by: s.malakhov s.malakhov@partner.samsung.com

[Torrero]

LGTM

[mhs4670go]

LGTM