Incorrectly judging the request as a bad request

[lvhan028]

Test the following cases

• case 1

# server
lmdeploy serve api_server
# client
v1/chat/completions + logprobs(true) + top_logprobs(0)
v1/chat/completions + logprobs(true) + top_logprobs(>=1)
v1/chat/completions + logprobs(false) + top_logprobs(0)
v1/chat/completions + logprobs(false) + top_logprobs(>=1)
v1/completions + logprobs(0)
v1/completions + logprobs(>0)
v1/completions + logprobs(<0)
/generate + return_logprobs(true)
/generate + return_logprobs(false)

• case 2

# server
lmdeploy serve api_server -logprobs-mode raw_logprobs
# client
v1/chat/completions + logprobs(true) + top_logprobs(0)
v1/chat/completions + logprobs(true) + top_logprobs(>=1)
v1/chat/completions + logprobs(false) + top_logprobs(0)
v1/chat/completions + logprobs(false) + top_logprobs(>=1)
v1/completions + logprobs(0)
v1/completions + logprobs(>0)
v1/completions + logprobs(<0)
/generate + return_logprobs(true)
/generate + return_logprobs(false)

[lvhan028]

cc @grimoire
When serving the model by enabling logprobs_mode, if users set logprobs=0 in a request, pytorch engine crashed since logprobs.vals is None