Fix repetition penalty for long context

[irexyc]

Motivation

address #1009 #898

去掉Shared Memory 支持更长的 session

[irexyc]

[irexyc]

before & after （llama-7b）

concurrency: 64
elapsed_time: 179.799s

first token latency(s)(min, max, ave): 0.031, 1.292, 0.106
per-token latency(s) percentile(50, 75, 95, 99): [0.024, 0.035, 0.053, 0.074]

number of prompt tokens: 490772
number of completion tokens: 472162
token throughput (completion token): 2626.051 token/s
token throughput (prompt + completion token): 5355.606 token/s
RPS (request per second): 11.124 req/s
RPM (request per minute): 667.411 req/min
--------------------------------------------------


concurrency: 64
elapsed_time: 179.440s

first token latency(s)(min, max, ave): 0.037, 1.361, 0.104
per-token latency(s) percentile(50, 75, 95, 99): [0.024, 0.033, 0.05, 0.073]

number of prompt tokens: 490772
number of completion tokens: 472162
token throughput (completion token): 2631.309 token/s
token throughput (prompt + completion token): 5366.330 token/s
RPS (request per second): 11.146 req/s
RPM (request per minute): 668.747 req/min
--------------------------------------------------

[lzhangzz]

before & after （llama-7b）

It seems that llama-7b does not use repetition penalty by default.

[irexyc]

It seems that llama-7b does not use repetition penalty by default.

I added repetition_penalty=1.005 in stream_infer manually.
https://github.com/InternLM/lmdeploy/blob/main/benchmark/profile_throughput.py#L88-L97