[WWB] Add eagle3 pipeline

[sunxiaoxia2022]

Description

Add eagle3 pipeline

Ticket: CVS-170888

Checklist:

• Tests have been updated or added to cover the new code
• This patch fully addresses the ticket.
• I have made corresponding changes to the documentation

[Copilot]

Pull Request Overview

This PR adds eagle3 pipeline support for speculative decoding in the who_what_benchmark tool. The changes enable users to configure and use draft models for speculative decoding with various configuration options.

• Added command-line arguments for speculative decoding configuration including draft model path, device, and eagle3 mode
• Modified text generation functions to use a unified generation config object instead of individual parameters
• Updated model loader to support draft model configuration and speculative decoding setup

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 4 comments.

File	Description
wwb.py	Added CLI arguments for speculative decoding and eagle3 mode, updated generation config handling
text_evaluator.py	Modified generation function signatures to use generation config object
model_loaders.py	Added draft model loading and configuration support for speculative decoding

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[sbalandi]

please, import openvino_genai and create GenerationConfig only if --genai option is set

[sbalandi]

you can create and set generation config once when you create the GenAI pipeline in model_loaders.py

[sunxiaoxia2022]

Ok, Updated.

[Copilot]

Pull Request Overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[MaximProshin]

@sunxiaoxia2022 , please share wwb Similarity numbers for eagle3 models from #2740

[sunxiaoxia2022]

@sunxiaoxia2022 , please share wwb Similarity numbers for eagle3 models from #2740

Hi @MaximProshin
Test platform: LNL Ultra 7 258V Win
Models:

1. llama-3.1-8b-instruct:
   target model: meta-llama/Llama-3.1-8B-Instruct
   Eagle3 draft model: yuhuili/EAGLE3-LLaMA3.1-Instruct-8B
2. qwen3-8b:
   target model: Qwen/Qwen3-8B
   Eagle3 draft model: Tengyunw/qwen3_8b_eagle3

The similarity numbers are as follows:

<style> </style>

model	precision	prompt	base	eagle pipeline
llama-3.1-8b-instruct	INT4	short	0.935972	0.935842
llama-3.1-8b-instruct	INT4	long	0.923789	0.918256
qwen3-8b	INT4	short	0.935486	0.935193
qwen3-8b	INT4	long	0.913537	0.914053

[apaniukov]

@sunxiaoxia2022 , please share wwb Similarity numbers for eagle3 models from #2740

Hi @MaximProshin Test platform: LNL Ultra 7 258V Win Models:

1. llama-3.1-8b-instruct:
   target model: meta-llama/Llama-3.1-8B-Instruct
   Eagle3 draft model: yuhuili/EAGLE3-LLaMA3.1-Instruct-8B
2. qwen3-8b:
   target model: Qwen/Qwen3-8B
   Eagle3 draft model: Tengyunw/qwen3_8b_eagle3

The similarity numbers are as follows:

<style> </style>

model precision prompt base eagle pipeline
llama-3.1-8b-instruct INT4 short 0.935972 0.935842
llama-3.1-8b-instruct INT4 long 0.923789 0.918256
qwen3-8b INT4 short 0.935486 0.935193
qwen3-8b INT4 long 0.913537 0.914053

What are num-assistant-tokens and assistant-confidence-threshold?

[Wovchena]

Why doesn't the similarity match exactly between base and eagle pipelines? I thought the expert model should be exactly the same

[sbalandi]

LGTM, as part with speculative decoding in wwb part, but still need clarification about Similarity numbers