happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

UI: add show/hide stats control to conversation screen's assistant message bubble; fix placeholder

This commit is contained in:
Han Yin 2025-08-29 18:56:43 -07:00
parent 8bd9615e6b
commit 027c68db64
5 changed files with 152 additions and 30 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
examples/llama.android/app/src/main/java/com/example/llama/engine/InferenceServices.kt
View File

@ -107,6 +107,7 @@ data class TokenMetrics(
val tokensCount: Int,
val ttftMs: Long,
val tpsMs: Float,
val duration: Long,
) {
val text: String
get() = "Tokens: $tokensCount, TTFT: ${ttftMs}ms, TPS: ${"%.1f".format(tpsMs)}"
@ -270,7 +271,8 @@ internal class InferenceServiceImpl @Inject internal constructor(
return TokenMetrics(
tokensCount = tokenCount,
ttftMs = if (firstTokenTime > 0) firstTokenTime - generationStartTime else 0L,
tpsMs = calculateTPS(tokenCount, totalTimeMs)
tpsMs = calculateTPS(tokenCount, totalTimeMs),
duration = totalTimeMs,
)
}

2
examples/llama.android/app/src/main/java/com/example/llama/ui/scaffold/bottombar/ConversationBottomBar.kt
View File

@ -45,7 +45,7 @@ fun ConversationBottomBar(
onAttachFileClick: (() -> Unit)?,
onAudioInputClick: (() -> Unit)?,
) {
val placeholder = if (isReady) "Message ${APP_NAME}..." else "Please wait for ${APP_NAME} to finish"
val placeholder = if (isReady) "Type your message here" else "AI is generating the response..."
Column(
modifier = Modifier.Companion.fillMaxWidth()

124
examples/llama.android/app/src/main/java/com/example/llama/ui/screens/ConversationScreen.kt
View File

@ -27,11 +27,16 @@ import androidx.compose.foundation.lazy.items
import androidx.compose.foundation.lazy.rememberLazyListState
import androidx.compose.foundation.shape.CircleShape
import androidx.compose.foundation.shape.RoundedCornerShape
import androidx.compose.material.icons.Icons
import androidx.compose.material.icons.filled.Timer
import androidx.compose.material3.Card
import androidx.compose.material3.CardDefaults
import androidx.compose.material3.ExperimentalMaterial3ExpressiveApi
import androidx.compose.material3.HorizontalDivider
import androidx.compose.material3.Icon
import androidx.compose.material3.MaterialTheme
import androidx.compose.material3.Text
import androidx.compose.material3.TonalToggleButton
import androidx.compose.runtime.Composable
import androidx.compose.runtime.DisposableEffect
import androidx.compose.runtime.LaunchedEffect
@ -55,18 +60,22 @@ import androidx.lifecycle.LifecycleEventObserver
import androidx.lifecycle.compose.LocalLifecycleOwner
import com.example.llama.data.model.ModelInfo
import com.example.llama.engine.ModelLoadingMetrics
import com.example.llama.engine.TokenMetrics
import com.example.llama.ui.components.ModelCardContentArchitectureRow
import com.example.llama.ui.components.ModelCardContentContextRow
import com.example.llama.ui.components.ModelCardContentField
import com.example.llama.ui.components.ModelCardCoreExpandable
import com.example.llama.ui.components.ModelUnloadDialogHandler
import com.example.llama.util.formatMilliSeconds
import com.example.llama.util.formatMilliSecondstructured
import com.example.llama.util.toEnglishName
import com.example.llama.viewmodel.ConversationViewModel
import com.example.llama.viewmodel.Message
import kotlinx.coroutines.FlowPreview
import kotlinx.coroutines.flow.distinctUntilChanged
import kotlinx.coroutines.flow.filter
import kotlinx.coroutines.launch
import java.util.Locale
private const val AUTO_REPOSITION_INTERVAL = 150L
@ -311,7 +320,7 @@ private fun MessageBubble(message: Message) {
content = message.content,
isThinking = false,
isGenerating = false,
metrics = message.metrics.text
metrics = message.metrics
)
}
}
@ -319,9 +328,7 @@ private fun MessageBubble(message: Message) {
@Composable
private fun UserMessageBubble(content: String, formattedTime: String) {
Column(
modifier = Modifier
.fillMaxWidth()
.padding(vertical = 8.dp),
modifier = Modifier.fillMaxWidth().padding(vertical = 8.dp),
horizontalAlignment = Alignment.End
) {
// Timestamp above bubble
@ -332,11 +339,12 @@ private fun UserMessageBubble(content: String, formattedTime: String) {
modifier = Modifier.padding(bottom = 4.dp)
)
Row {
Row(modifier = Modifier.fillMaxWidth(0.9f)) {
Spacer(modifier = Modifier.weight(1f))
Card(
shape = RoundedCornerShape(16.dp, 4.dp, 16.dp, 16.dp),
modifier = Modifier,
shape = RoundedCornerShape(16.dp, 2.dp, 16.dp, 16.dp),
colors = CardDefaults.cardColors(
containerColor = MaterialTheme.colorScheme.primaryContainer
),
@ -359,7 +367,7 @@ private fun AssistantMessageBubble(
content: String,
isThinking: Boolean,
isGenerating: Boolean,
metrics: String? = null
metrics: TokenMetrics? = null
) {
Row(
verticalAlignment = Alignment.Top,
@ -381,10 +389,7 @@ private fun AssistantMessageBubble(
Spacer(modifier = Modifier.width(8.dp))
Column(
modifier = Modifier
.fillMaxWidth()
) {
Column(modifier = Modifier.fillMaxWidth(0.9f)) {
// Timestamp above bubble
if (formattedTime.isNotBlank()) {
Text(
@ -396,7 +401,7 @@ private fun AssistantMessageBubble(
}
Card(
shape = RoundedCornerShape(4.dp, 16.dp, 16.dp, 16.dp),
shape = RoundedCornerShape(2.dp, 16.dp, 16.dp, 16.dp),
colors = CardDefaults.cardColors(
containerColor = MaterialTheme.colorScheme.surfaceVariant
),
@ -412,13 +417,11 @@ private fun AssistantMessageBubble(
}
// Show metrics or generation status below the bubble
if (isGenerating) {
Row(
modifier = Modifier
.height(20.dp)
.padding(top = 4.dp),
modifier = Modifier.height(20.dp).padding(top = 4.dp),
verticalAlignment = Alignment.CenterVertically
) {
if (isGenerating) {
PulsatingDots(small = true)
Spacer(modifier = Modifier.width(4.dp))
@ -428,13 +431,11 @@ private fun AssistantMessageBubble(
style = MaterialTheme.typography.labelSmall,
color = MaterialTheme.colorScheme.primary
)
} else if (metrics != null) {
}
} else {
// Show metrics when message is complete
Text(
text = metrics,
style = MaterialTheme.typography.labelSmall,
color = MaterialTheme.colorScheme.onSurfaceVariant.copy(alpha = 0.7f),
)
metrics?.let {
ExpandableTokenMetricsBubble(metrics)
}
}
}
@ -480,3 +481,82 @@ private fun PulsatingDots(small: Boolean = false) {
}
}
}
@OptIn(ExperimentalMaterial3ExpressiveApi::class)
@Composable
private fun ExpandableTokenMetricsBubble(metrics: TokenMetrics) {
var showMetrics by remember { mutableStateOf(false) }
Column {
TonalToggleButton(
checked = showMetrics,
onCheckedChange = { showMetrics = !showMetrics }
) {
Icon(
modifier = Modifier.size(16.dp),
imageVector = Icons.Default.Timer,
contentDescription = "${if (showMetrics) "Hide" else "Show"} token metrics of this assistant message"
)
Text(
text = "${if (showMetrics) "Hide" else "Show"} stats",
modifier = Modifier.padding(start = 8.dp),
style = MaterialTheme.typography.labelMedium,
)
}
if (showMetrics) {
Card(
shape = RoundedCornerShape(2.dp, 16.dp, 16.dp, 16.dp),
colors = CardDefaults.cardColors(
containerColor = MaterialTheme.colorScheme.surfaceVariant
),
elevation = CardDefaults.cardElevation(defaultElevation = 1.dp)
) {
Row(
modifier = Modifier.padding(12.dp),
horizontalArrangement = Arrangement.spacedBy(18.dp)
) {
val ttft = formatMilliSecondstructured(metrics.ttftMs)
TokenMetricSection("1st Token", ttft.value.toString(), ttft.unit.toEnglishName())
val tps = String.format(Locale.getDefault(), "%.2f", metrics.tpsMs)
TokenMetricSection("Decode speed", tps, "tokens/sec")
val duration = formatMilliSecondstructured(metrics.duration)
TokenMetricSection("Duration", duration.value.toString(), duration.unit.toEnglishName())
}
}
}
}
}
@Composable
private fun TokenMetricSection(metricName: String, metricValue: String, metricUnit: String) {
Column {
Text(
text = metricName,
style = MaterialTheme.typography.labelSmall,
color = MaterialTheme.colorScheme.onSurfaceVariant,
fontWeight = FontWeight.Normal,
)
Text(
modifier = Modifier.padding(top = 2.dp),
text = metricValue,
style = MaterialTheme.typography.bodyLarge,
color = MaterialTheme.colorScheme.onSurfaceVariant,
fontWeight = FontWeight.Light,
fontStyle = FontStyle.Italic
)
Text(
modifier = Modifier.padding(top = 2.dp),
text = metricUnit,
style = MaterialTheme.typography.labelSmall,
color = MaterialTheme.colorScheme.onSurfaceVariant.copy(alpha = 0.7f),
fontWeight = FontWeight.Light,
fontStyle = FontStyle.Italic
)
}
}

40
examples/llama.android/app/src/main/java/com/example/llama/util/FormatUtils.kt
View File

@ -1,6 +1,21 @@
package com.example.llama.util
import java.util.concurrent.TimeUnit
import java.util.Locale
import kotlin.math.round
/**
* Maps [TimeUnit] to English names, with plural form support
*/
fun TimeUnit.toEnglishName(plural: Boolean = false): String = when (this) {
TimeUnit.NANOSECONDS -> if (plural) "nanoseconds" else "nanosecond"
TimeUnit.MICROSECONDS -> if (plural) "microseconds" else "microsecond"
TimeUnit.MILLISECONDS -> if (plural) "milliseconds" else "millisecond"
TimeUnit.SECONDS -> if (plural) "seconds" else "second"
TimeUnit.MINUTES -> if (plural) "minutes" else "minute"
TimeUnit.HOURS -> if (plural) "hours" else "hour"
TimeUnit.DAYS -> if (plural) "days" else "day"
}
/**
* Formats milliseconds into a human-readable time string
@ -18,6 +33,31 @@ fun formatMilliSeconds(millis: Long): String {
}
}
data class DurationValue(
val value: Double,
val unit: TimeUnit
)
/**
* Converts milliseconds into a structured DurationValue.
*
* Rules:
* - < 100 seconds -> show in SECONDS
* - < 100 minutes -> show in MINUTES
* - < 100 hours -> show in HOURS
*/
fun formatMilliSecondstructured(millis: Long): DurationValue {
val seconds = millis / 1000.0
return when {
seconds < 100 -> DurationValue(round2(seconds), TimeUnit.SECONDS)
seconds < 100 * 60 -> DurationValue(round2(seconds / 60.0), TimeUnit.MINUTES)
else -> DurationValue(round2(seconds / 3600.0), TimeUnit.HOURS)
}
}
private fun round2(v: Double): Double = round(v * 100) / 100
/**
* Convert bytes into human readable sizes
*/

4
examples/llama.android/app/src/main/java/com/example/llama/viewmodel/ConversationViewModel.kt
View File

@ -62,14 +62,14 @@ class ConversationViewModel @Inject constructor(
content = content,
timestamp = System.currentTimeMillis()
)
_messages.value = _messages.value + userMessage
_messages.value += userMessage
// Add placeholder for assistant response
val assistantMessage = Message.Assistant.Ongoing(
content = "",
timestamp = System.currentTimeMillis()
)
_messages.value = _messages.value + assistantMessage
_messages.value += assistantMessage
// Clear input field
inputFieldState.clearText()