[AWQ] Remove Code to obtain Merge Node

[anzr299]

Changes

Remove Code to obtain Merge Node. Instead directly get the merge node from the match nodes.

Reason for changes

Matmul1 -> mul -> Matmul2 pattern was causing an error when the precision for Matmul1 was FP32. The first matmul node was not being found in all_weight_params.

According to my experiments, in both the conditions for patterns MatMul->Multiply->MatMul and Act->MatMul or Act->Multiply->MatMul we essentially do the same task of obtaining the first node from the matched list of nodes.

Tests

Category	Job	Status	Job Number	Notes
GH Examples	Examples	Pass	382
GH WC	Weights Compression	Pass	137

Perplexity Evaluation:
Compression Settings: Ratio=0.8, Hessian Input Activation, backup_mode NONE
Evaluation Settings: limit 10(evaluate with only 10 samples in lm_eval)
Model: microsoft/Phi-3-mini-128k-instruct

Branch	AWQ	Accuracy
Develop Branch	No	11.157
Develop Branch	Yes	11.0583
After Change	Yes	11.0583

Related tickets

172326

[daniil-lyakhov]

@anzr299, please check #2708 and ticket 141131. Looks like this code was added to support some LLMs, we need to make sure that this models are compressed as expected

[anzr299]

@anzr299, please check #2708 and ticket 141131. Looks like this code was added to support some LLMs, we need to make sure that this models are compressed as expected

Ah I see, I will apply another workaround instead.

[anzr299]

@anzr299, please check #2708 and ticket 141131. Looks like this code was added to support some LLMs, we need to make sure that this models are compressed as expected

I have verified by checking the accuracy of the microsoft/Phi-3-mini-128k-instruct model before and after the change with AWQ enabled. It is the same. Looks like this change does not affect the model.

[andreyanufr]

Checked on llama (MatMul->Multiply->MatMul) and phi (activations->MatMul). The new changes do not affect the insertion location of the scale.