support qwen3-next eagle3

[sleepcoo]

Motivation

support qwen3 next eagle3,https://huggingface.co/lukeysong/qwen3-next-draft/tree/main

Benchmarking and Profiling

use Specfoge

python bench_eagle3.py --model /workdir/huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct-FP8 --speculative-algorithm EAGLE3 --speculative-draft-model-path /workdir/epoch_1_step_120000/ --port 30000 --config-list 1,3,1,4 --dtype bfloat16 --tp 4

tp4 h20

Benchmark	Latency (ms)	Throughput (tok/s)	Accept Length	Accuracy	Num Questions	Valid Predictions	Num Samples
mtbench	461.65	234.52	2.59	N/A	80	0	80
gsm8k	137.10	234.18	3.13	0.955	200	200	200
humaneval	309.11	304.63	3.45	0.000	164	164	200
math500	515.99	317.04	3.52	0.615	200	200	200

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[sleepcoo]

/tag-and-rerun-ci