mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 33 Packages Projects Releases Wiki Activity

server, webui: support continue generation on reasoning models (#22727)

Browse Source 
* server, webui : support continue generation on reasoning models (#22727)

Remove the throw blocking assistant prefill on reasoning models and
orchestrate thinking tags around the prefilled message so the parser
routes the next stream chunks correctly. WebUI drops the reasoning
guard on the Continue button, sends reasoning_content with the
prefilled message and persists partial reasoning on stop so the CoT
survives reload and resume.

Scope : templates with a simple thinking_start_tag / thinking_end_tag
pair. Channel-based templates like GPT-OSS are out of scope, pending
a per-template prefill API in common/chat.

First step toward #21754.

* chore: update webui build output

* server: reject reasoning prefill on channel based templates
This commit is contained in:
Pascal
2026-05-13 11:09:51 +02:00
committed by GitHub
parent 3796c94bad
commit 5d44db6008
5 changed files with 250 additions and 193 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessageAssistant/ChatMessageAssistant.svelte
+1 -2
View File
@@ -74,7 +74,6 @@
const editCtx = getMessageEditContext();
const isAgentic = $derived(hasAgenticContent(message, toolMessages));
const hasReasoning = $derived(!!message.reasoningContent);
const processingState = useProcessingState();
let currentConfig = $derived(config());
@@ -329,7 +328,7 @@
{onCopy}
{onEdit}
{onRegenerate}
onContinue={currentConfig.enableContinueGeneration && !hasReasoning ? onContinue : undefined}
onContinue={currentConfig.enableContinueGeneration ? onContinue : undefined}
{onForkConversation}
{onDelete}
{onConfirmDelete}
tools/server/webui/src/lib/constants/settings-registry.ts
+1 -1
View File
@@ -122,7 +122,7 @@ const SETTINGS_REGISTRY: Record<string, SettingsSectionEntry> = {
{
key: SETTINGS_KEYS.ENABLE_CONTINUE_GENERATION,
label: 'Enable "Continue" button',
help: 'Enable "Continue" button for assistant messages. Currently works only with non-reasoning models.',
help: 'Enable "Continue" button for assistant messages, including reasoning models.',
defaultValue: false,
type: SettingsFieldType.CHECKBOX,
section: SETTINGS_SECTION_SLUGS.GENERAL,
tools/server/webui/src/lib/stores/chat.svelte.ts
+47 -27
View File
@@ -674,7 +674,8 @@ class ChatStore {
},
onReasoningChunk: (chunk: string) => {
streamedReasoningContent += chunk;
// Update UI to show reasoning is being received
// mark streaming state so a stop mid-thinking can persist the partial reasoning
this.setChatStreaming(convId, streamedContent, currentMessageId);
const idx = conversationsStore.findMessageIndex(currentMessageId);
conversationsStore.updateMessageAtIndex(idx, {
reasoningContent: streamedReasoningContent
@@ -989,38 +990,51 @@ class ChatStore {
const conversationId = convId || conversationsStore.activeConversation?.id;
if (!conversationId) return;
const streamingState = this.getChatStreaming(conversationId);
if (!streamingState || !streamingState.response.trim()) return;
if (!streamingState) return;
const messages =
conversationId === conversationsStore.activeConversation?.id
? conversationsStore.activeMessages
: await conversationsStore.getConversationMessages(conversationId);
if (!messages.length) return;
const lastMessage = messages[messages.length - 1];
if (lastMessage?.role === MessageRole.ASSISTANT) {
try {
const updateData: { content: string; timings?: ChatMessageTimings } = {
content: streamingState.response
};
const lastKnownState = this.getProcessingState(conversationId);
if (lastKnownState) {
updateData.timings = {
prompt_n: lastKnownState.promptTokens || 0,
prompt_ms: lastKnownState.promptMs,
predicted_n: lastKnownState.tokensDecoded || 0,
cache_n: lastKnownState.cacheTokens || 0,
predicted_ms:
lastKnownState.tokensPerSecond && lastKnownState.tokensDecoded
? (lastKnownState.tokensDecoded / lastKnownState.tokensPerSecond) * 1000
: undefined
};
}
await DatabaseService.updateMessage(lastMessage.id, updateData);
lastMessage.content = streamingState.response;
if (updateData.timings) lastMessage.timings = updateData.timings;
} catch (error) {
lastMessage.content = streamingState.response;
console.error('Failed to save partial response:', error);
if (lastMessage?.role !== MessageRole.ASSISTANT) return;
const partialContent = streamingState.response;
const partialReasoning = lastMessage.reasoningContent || '';
// nothing to persist when both content and reasoning are empty (e.g. stop before any token)
if (!partialContent.trim() && !partialReasoning.trim()) return;
try {
const updateData: {
content: string;
reasoningContent?: string;
timings?: ChatMessageTimings;
} = {
content: partialContent
};
if (partialReasoning) {
updateData.reasoningContent = partialReasoning;
}
const lastKnownState = this.getProcessingState(conversationId);
if (lastKnownState) {
updateData.timings = {
prompt_n: lastKnownState.promptTokens || 0,
prompt_ms: lastKnownState.promptMs,
predicted_n: lastKnownState.tokensDecoded || 0,
cache_n: lastKnownState.cacheTokens || 0,
predicted_ms:
lastKnownState.tokensPerSecond && lastKnownState.tokensDecoded
? (lastKnownState.tokensDecoded / lastKnownState.tokensPerSecond) * 1000
: undefined
};
}
await DatabaseService.updateMessage(lastMessage.id, updateData);
lastMessage.content = partialContent;
if (updateData.timings) lastMessage.timings = updateData.timings;
} catch (error) {
lastMessage.content = partialContent;
console.error('Failed to save partial response:', error);
}
}
@@ -1265,7 +1279,11 @@ class ChatStore {
const conversationContext = conversationsStore.activeMessages.slice(0, idx);
const contextWithContinue = [
...conversationContext,
{ role: MessageRole.ASSISTANT as const, content: originalContent }
{
role: MessageRole.ASSISTANT as const,
content: originalContent,
reasoning_content: originalReasoning || undefined
}
];
let appendedContent = '';
@@ -1291,6 +1309,8 @@ class ChatStore {
onReasoningChunk: (chunk: string) => {
appendedReasoning += chunk;
hasReceivedContent = true;
// mark streaming state so a stop mid-thinking can persist the partial reasoning
this.setChatStreaming(msg.convId, originalContent + appendedContent, msg.id);
conversationsStore.updateMessageAtIndex(idx, {
reasoningContent: originalReasoning + appendedReasoning
});