重复输出 about deepseek-coder

[guoday]

请问是什么规模的模型(1b,7b,33b)，用了什么类型的模型(base, instruct)，大概补全了多少行

[gafield-liu]

33b base, 6.7b base模型，case是50行左右。

[gafield-liu]

测了一些例子，6.7b模型大概有30%的概率，会输出不停～

[guoday]

输出不停是指无法遇到还是说会重复输出同样的代码？

[guoday]

你可以试下Repetition penalty设为1.1，我们在使用补全插件的过程中，Repetition penalty设为1.1，体验会好一些

[gafield-liu]

类似这个case
<｜fim▁begin｜>
package com.vitorpamplona.amethyst

import android.app.PendingIntent
import android.content.Context
import android.content.Intent
import android.util.LruCache
import androidx.core.net.toUri
import androidx.media3.common.C
import androidx.media3.common.Player
import androidx.media3.common.Player.STATE_IDLE
import androidx.media3.common.Player.STATE_READY
import androidx.media3.exoplayer.ExoPlayer
import androidx.media3.session.MediaSession
import com.vitorpamplona.amethyst.ui.MainActivity
import kotlinx.coroutines.DelicateCoroutinesApi
import kotlinx.coroutines.Dispatchers
import kotlinx.coroutines.GlobalScope
import kotlinx.coroutines.launch
import kotlin.math.abs

class MultiPlayerPlaybackManager(
private val dataSourceFactory: androidx.media3.exoplayer.source.MediaSource.Factory? = null,
private val cachedPositions: VideoViewedPositionCache
) {
// protects from LruCache killing playing sessions
private val playingMap = mutableMapOf<String, MediaSession>()

private val cache =
object : LruCache<String, MediaSession>(10) { // up to 10 videos in the screen at the same time
override fun entryRemoved(
evicted: Boolean,
key: String?,
oldValue: MediaSession?,
newValue: MediaSession?
) {
<｜fim▁hole｜>
}
}

private fun getCallbackIntent(callbackUri: String, applicationContext: Context): PendingIntent {
return PendingIntent.getActivity(
applicationContext,
0,
Intent(Intent.ACTION_VIEW, callbackUri.toUri(), applicationContext, MainActivity::class.java),
PendingIntent.FLAG_IMMUTABLE or PendingIntent.FLAG_UPDATE_CURRENT
)
<｜fim▁end｜> if (evicted) {
oldValue?.release()
}
}
}

fun getSession(
videoUrl: String,
applicationContext: Context,
callbackUri: String,
onPlayerReady: (MediaSession) -> Unit
): MediaSession {
val session = cache.get(videoUrl)
if (session != null) {
return session
}

val newSession = MediaSession(applicationContext, videoUrl)
newSession.setCallback(object : MediaSession.Callback() {
    override fun onPlay() {
        super.onPlay()
        playingMap[videoUrl] = newSession
    }

    override fun onPause() {
        super.onPause()
        playingMap.remove(videoUrl)
    }
})

newSession.setMediaButtonEventReceiver(getCallbackIntent(callbackUri, applicationContext))

newSession.setPlaybackToLocal()
newSession.setRatingType(C.RATING_TYPE_NONE)
newSession.setPlayWhenReady(true)

val player = ExoPlayer.Builder(applicationContext).build()
player.addListener(object : Player.Listener {
    override fun onPlaybackStateChanged(playbackState: Int) {
        if (playbackState == STATE_READY) {
            onPlayerReady(newSession)
        }
    }
})

newSession.setPlayer(player)
newSession.setMediaSource(dataSourceFactory?.createMediaSource(videoUrl.toUri()))

cache.put(videoUrl, newSession)

return newSession

}

fun release(videoUrl: String) {
cache.remove(videoUrl)
}

fun releaseAll() {
cache.evictAll()
}

fun getPlayingSession(): MediaSession? {
return playingMap.values.firstOrNull()
}

fun getPlayingSession(videoUrl: String): MediaSession? {
return playingMap[videoUrl]
}

fun getPlayingSession(videoUrl: String, onPlayerReady: (MediaSession) -> Unit) {
val session = playingMap[videoUrl]
if (session != null) {
onPlayerReady(session)
}
}

fun getPlayingSession(videoUrl: String, onPlayerReady: (MediaSession) -> Unit, onPlayerNotReady: () -> Unit) {
val session = playingMap[videoUrl]
if (session != null) {
onPlayerReady(session)
} else {
onPlayerNotReady()
}
}

fun getPlayingSession(videoUrl: String, onPlayerReady: (MediaSession) -> Unit, onPlayerNotReady: () -> Unit, onPlayerError: () -> Unit) {
val session = playingMap[videoUrl]
if (session != null) {
onPlayerReady(session)
} else {
onPlayerNotReady()
}
}

fun getPlayingSession(videoUrl: String, onPlayerReady: (MediaSession) -> Unit, onPlayerNotReady: () -> Unit, onPlayerError: () -> Unit, onPlayerIdle: () -> Unit) {
val session = playingMap[videoUrl]
if (session != null) {
onPlayerReady(session)
} else {
onPlayerNotReady()
}
}

fun getPlayingSession(videoUrl: String, onPlayerReady: (MediaSession) -> Unit, onPlayerNotReady: () -> Unit, onPlayerError: () -> Unit, onPlayerIdle: () -> Unit, onPlayerEnded: () -> Unit) {
val session = playingMap[videoUrl]
if (session != null) {
onPlayerReady(session)
} else {
onPlayerNotReady()
}
}

fun getPlayingSession(videoUrl: String, onPlayerReady: (MediaSession) -> Unit, onPlayerNotReady: () -> Unit, onPlayerError: () -> Unit, onPlayerIdle: () -> Unit, onPlayerEnded: () -> Unit, onPlayerBuffering:

[pkuzqh]

训练的时候fim确实会随机剪开文件拼在一起，因此有时候会出现补全比较长的情况，如果是使用成补全的话，最好设置最大长度

[gafield-liu]

嗯嗯明白，似乎没找到一种合适的配置，可以让模型补全的比较稳定～
试过了codellama的补全功能也有类似的情况。

[guoday]

我们一般使用fim的话，会用Repetition penalty为1，然后限制补全最多3行或遇到eos。如果要继续补全，就再过一次fim

[gafield-liu]

收到，感谢回复～