spec: add backend sampling support for eagle3

[ruixiang63]

Overview

Following #23287 to add backend sampling support for eagle3.

Performance results on SpeedBench

• eagle3 baseline

python tools/server/bench/speed-bench/speed_bench.py --url localhost:8080 --bench qualitative --category all --osl 512 --concurrency 1 --limit 5 --output eagle3-qwen3-baseline.json

Summary (elapsed=353.27s)
category       samples  avg_prompt_t/s  avg_pred_t/s  avg_latency  accept_rate
-------------  -------  --------------  ------------  -----------  -----------
coding         5        1804.50         96.32         6.580s       0.4810     
humanities     5        446.18          82.61         7.500s       0.3669     
math           5        46.57           82.13         6.273s       0.3605     
qa             5        605.86          92.40         4.808s       0.4444     
rag            5        4787.48         97.18         6.307s       0.5010     
reasoning      5        201.42          82.05         6.298s       0.3574     
stem           5        47.51           82.29         6.259s       0.3605     
writing        5        4234.16         91.77         6.931s       0.4415     
multilingual   5        2644.99         95.54         5.463s       0.4741     
summarization  5        628.52          87.96         3.667s       0.4104     
roleplay       5        2782.64         89.17         10.565s      0.4230     
overall        55       1657.26         89.04         6.423s       0.4185     

• eagle3 with backend sampling

python tools/server/bench/speed-bench/speed_bench.py --url localhost:8080 --bench qualitative --category all --osl 512 --concurrency 1 --limit 5 --output eagle3-qwen3-backend-sampling.json

Summary (elapsed=351.79s)
category       samples  avg_prompt_t/s  avg_pred_t/s  avg_latency  accept_rate
-------------  -------  --------------  ------------  -----------  -----------
coding         5        1817.48         97.18         6.577s       0.4810     
humanities     5        455.30          83.50         7.414s       0.3669     
math           5        46.17           83.14         6.196s       0.3605     
qa             5        622.08          93.79         4.736s       0.4444     
rag            5        3754.37         94.52         6.738s       0.5010     
reasoning      5        202.27          82.78         6.243s       0.3574     
stem           5        47.31           83.19         6.191s       0.3605     
writing        5        4273.63         93.15         6.835s       0.4415     
multilingual   5        2613.75         96.83         5.393s       0.4741     
summarization  5        642.39          89.23         3.602s       0.4104     
roleplay       5        2700.37         90.38         10.431s      0.4230     
overall        55       1561.37         89.79         6.396s       0.4185 

• comparison

python tools/server/bench/speed-bench/speed_bench_compare.py --baseline eagle3-qwen3-baseline.json --speculative eagle3-qwen3-backend-sampling.json

Comparison: baseline=eagle3-qwen3-baseline.json speculative=eagle3-qwen3-backend-sampling.json
category       base_avg_pred_t/s  spec_avg_pred_t/s  decode_speedup  base_avg_latency  spec_avg_latency  latency_speedup  accept_rate
-------------  -----------------  -----------------  --------------  ----------------  ----------------  ---------------  -----------
coding         96.32              95.56              0.99x           6.580s            6.668s            0.99x            0.4810     
humanities     82.61              83.96              1.02x           7.500s            7.381s            1.02x            0.3669     
math           82.13              83.57              1.02x           6.273s            6.163s            1.02x            0.3605     
qa             92.40              93.98              1.02x           4.808s            4.725s            1.02x            0.4444     
rag            97.18              98.35              1.01x           6.307s            6.329s            1.00x            0.5010     
reasoning      82.05              83.00              1.01x           6.298s            6.225s            1.01x            0.3574     
stem           82.29              83.26              1.01x           6.259s            6.187s            1.01x            0.3605     
writing        91.77              92.99              1.01x           6.931s            6.843s            1.01x            0.4415     
multilingual   95.54              96.79              1.01x           5.463s            5.431s            1.01x            0.4741     
summarization  87.96              89.06              1.01x           3.667s            3.612s            1.02x            0.4104     
roleplay       89.17              90.50              1.01x           10.565s           10.430s           1.01x            0.4230     
overall        89.04              90.09              1.01x           6.423s            6.363s            1.01x            0.4185    

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO