fix flash kernel on sm 70

[grimoire]

• flash attention triton kernel on sm 70 bug fixing
• profile throughput bug fixing

[lvhan028]

chat failed on v100

double enter to end input >>> hi, what's your name

<|im_start|>system
You are an AI assistant whose name is InternLM (书生·浦语).
- InternLM (书生·浦语) is a conversational language model that is developed by Shanghai AI Laboratory (上海人工智能实验室). It is designed to be helpful, honest, and harmless.
- InternLM (书生·浦语) can understand and communicate fluently in the language chosen by the user such as English and 中文.
<|im_end|>
<|im_start|>user
hi, what's your name<|im_end|>
<|im_start|>assistant
 S1 real

double enter to end input >>>

[grimoire]

chat failed on v100

can not reproduce with codellama.
Guess the error of internlm is caused by bf16, which is not supported on sm70.

[lvhan028]

I used internlm2-chat-7b, which is fp16

[grimoire]

Done