CodeQwen returns extra white space for code completion (w/server + fim)

[wsxiaoys]

Tested in b2724 and b2781 (latest)

start server

./server -m ~/.tabby/models/TabbyML/CodeQwen-7B/ggml/q8_0.v2.gguf -c 2048 --port 8081

curl --request POST \
  --url http://localhost:8081/completion \
  --header 'content-type: application/json' \
  --data '{"prompt": "<fim_prefix>// Path: crates/tabby/src/main.rs\n//     /// Download the language model for serving.\n//     Download(download::DownloadArgs),\n//\n// Path: crates/tabby/src/main.rs\n// #[derive(Subcommand)]\n// pub enum Commands {\n//     /// Starts the api endpoint for IDE / Editor extensions.\n//     Serve(serve::ServeArgs),\n//\n//     /// Download the language model for serving.\n//     Download(download::DownloadArgs),\n//\n//     /// Run scheduler progress for cron jobs integrating external code repositories.\n//     Scheduler(SchedulerArgs),\n//\n//     /// Run completion model as worker\n//     #[cfg(feature = \"ee\")]\n//     #[clap(name = \"worker::completion\", hide = true)]\n//     WorkerCompletion(worker::WorkerArgs),\n//\n        Commands::Scheduler(SchedulerArgs {\n            now,\n            url: Some(url),\n            token: Some(token),\n        }) => {\n            let client = tabby_webserver::public::create_scheduler_client(&url, &token).await;\n            tabby_scheduler::scheduler(now, client).await\n        }\n        Commands::Scheduler(SchedulerArgs { now, .. }) => {\n            tabby_scheduler::scheduler(now, ConfigRepositoryAccess).await\n        }\n        #[cfg(feature = \"ee\")]\n        Commands::WorkerCompletion(ref args) => {\n            worker::main(tabby_webserver::public::WorkerKind::Completion, args).await\n        }\n        #[cfg(feature = \"ee\")]\n        Commands::WorkerChat(ref args) => {\n            worker::main(tabby_webserver::public::WorkerKind::Chat, args).await\n        }\n        Commands::Download(<fim_suffix>)\n    }\n}\n\n#[macro_export]\nmacro_rules! fatal {\n    ($msg:expr) => {\n        ({\n            tracing::error!($msg);\n            std::process::exit(1);\n        })\n    };\n\n    ($fmt:expr, $($arg:tt)*) => {\n        ({\n            tracing::error!($fmt, $($arg)*);\n            std::process::exit(1);\n        })\n    };\n}\n<fim_middle>", "n_predict": 12}'

Output:

{
  "content": " ref args) => {\n            tabby_webs",
  ...
}

maybe related: The generated token has a _ prefix in HF vocab file.