Merge branch 'main' into dev

colab/GPU.ipynb

@@ -95,6 +95,7 @@
         "%cd /\n",
         "!git clone https://github.com/Cohee1207/SillyTavern-extras\n",
         "%cd /SillyTavern-extras\n",
+        "!git clone https://github.com/Cohee1207/tts_samples\n",
         "!npm install -g localtunnel\n",
         "!pip install -r requirements-complete.txt\n",
         "!pip install tensorflow==2.11\n",

public/script.js

@@ -1393,7 +1393,7 @@ function getExtensionPrompt(position = 0, depth = undefined, separator = "\n") {
 
 function baseChatReplace(value, name1, name2) {
     if (value !== undefined && value.length > 0) {
-        value = substituteParams(value, is_pygmalion ? "You:" : name1, name2);
+        value = substituteParams(value, is_pygmalion ? "You" : name1, name2);
 
         if (power_user.collapse_newlines) {
             value = collapseNewlines(value);
@@ -1410,9 +1410,10 @@ function appendToStoryString(value, prefix) {
 }
 
 function isStreamingEnabled() {
-    return (main_api == 'openai' && oai_settings.stream_openai)
+    return ((main_api == 'openai' && oai_settings.stream_openai)
         || (main_api == 'poe' && poe_settings.streaming)
-        || (main_api == 'textgenerationwebui' && textgenerationwebui_settings.streaming);
+        || (main_api == 'textgenerationwebui' && textgenerationwebui_settings.streaming))
+        && !isMultigenEnabled(); // Multigen has a quasi-streaming mode which breaks the real streaming
 }
 
 class StreamingProcessor {
@@ -1592,7 +1593,15 @@ async function Generate(type, { automatic_trigger, force_name2, resolve, reject,
 
     const isImpersonate = type == "impersonate";
     const isInstruct = power_user.instruct.enabled;
-    message_already_generated = isImpersonate ? `${name1}: ` : `${name2}: `;
+
+    // Name for the multigen prefix
+    const magName = isImpersonate ? (is_pygmalion ? 'You' : name1) : name2;
+
+    if (isInstruct) {
+        message_already_generated = formatInstructModePrompt(magName, isImpersonate);
+    } else {
+        message_already_generated = `${magName}: `;
+    }
 
     const interruptedByCommand = processCommands($("#send_textarea").val(), type);
 
@@ -2018,6 +2027,13 @@ async function Generate(type, { automatic_trigger, force_name2, resolve, reject,
                     const isBottom = j === mesSend.length - 1;
                     mesSendString += mesSend[j];
 
+                    // Add quiet generation prompt at depth 0
+                    if (isBottom && quiet_prompt && quiet_prompt.length) {
+                        const name = is_pygmalion ? 'You' : name1;
+                        const quietAppend = isInstruct ? formatInstructModeChat(name, quiet_prompt, true) : `\n${name}: ${quiet_prompt}`;
+                        mesSendString += quietAppend;
+                    }
+
                     if (isInstruct && isBottom && tokens_already_generated === 0) {
                         const name = isImpersonate ? (is_pygmalion ? 'You' : name1) : name2;
                         mesSendString += formatInstructModePrompt(name, isImpersonate);
@@ -2151,11 +2167,6 @@ async function Generate(type, { automatic_trigger, force_name2, resolve, reject,
                 }
             }
 
-            // Add quiet generation prompt at depth 0
-            if (quiet_prompt && quiet_prompt.length) {
-                finalPromt += `\n${quiet_prompt}`;
-            }
-
             finalPromt = finalPromt.replace(/\r/gm, '');
 
             if (power_user.collapse_newlines) {
@@ -2270,11 +2281,14 @@ async function Generate(type, { automatic_trigger, force_name2, resolve, reject,
                 hideSwipeButtons();
                 let getMessage = await streamingProcessor.generate();
 
+                // Cohee: Basically a dead-end code... (disabled by isStreamingEnabled)
+                // I wasn't able to get multigen working with real streaming
+                // consistently without screwing the interim prompting
                 if (isMultigenEnabled()) {
-                    tokens_already_generated += this_amount_gen;    // add new gen amt to any prev gen counter..
+                    tokens_already_generated += this_amount_gen;
                     message_already_generated += getMessage;
                     promptBias = '';
-                    if (!streamingProcessor.isStopped && shouldContinueMultigen(getMessage)) {
+                    if (!streamingProcessor.isStopped && shouldContinueMultigen(getMessage, isImpersonate)) {
                         streamingProcessor.isFinished = false;
                         runGenerate(getMessage);
                         console.log('returning to make generate again');
@@ -2306,16 +2320,23 @@ async function Generate(type, { automatic_trigger, force_name2, resolve, reject,
 
                         let this_mes_is_name;
                         ({ this_mes_is_name, getMessage } = extractNameFromMessage(getMessage, force_name2, isImpersonate));
-                        if (tokens_already_generated == 0) {
+
-                            console.log("New message");
+                        if (!isImpersonate) {
-                            ({ type, getMessage } = saveReply(type, getMessage, this_mes_is_name, title));
+                            if (tokens_already_generated == 0) {
-                        }
+                                console.log("New message");
-                        else {
+                                ({ type, getMessage } = saveReply(type, getMessage, this_mes_is_name, title));
-                            console.log("Should append message");
+                            }
-                            ({ type, getMessage } = saveReply('append', getMessage, this_mes_is_name, title));
+                            else {
+                                console.log("Should append message");
+                                ({ type, getMessage } = saveReply('append', getMessage, this_mes_is_name, title));
+                            }
+                        } else {
+                            let chunk = cleanUpMessage(message_already_generated, true);
+                            let extract = extractNameFromMessage(chunk, force_name2, isImpersonate);
+                            $('#send_textarea').val(extract.getMessage).trigger('input');
                         }
 
-                        if (shouldContinueMultigen(getMessage)) {
+                        if (shouldContinueMultigen(getMessage, isImpersonate)) {
                             hideSwipeButtons();
                             tokens_already_generated += this_amount_gen;            // add new gen amt to any prev gen counter..
                             getMessage = message_already_generated;
@@ -2335,6 +2356,7 @@ async function Generate(type, { automatic_trigger, force_name2, resolve, reject,
                     if (getMessage.length > 0) {
                         if (isImpersonate) {
                             $('#send_textarea').val(getMessage).trigger('input');
+                            generatedPromtCache = "";
                         }
                         else if (type == 'quiet') {
                             resolve(getMessage);
@@ -2624,12 +2646,24 @@ function getGenerateUrl() {
     return generate_url;
 }
 
-function shouldContinueMultigen(getMessage) {
+function shouldContinueMultigen(getMessage, isImpersonate) {
-    const nameString = is_pygmalion ? 'You:' : `${name1}:`;
+    if (power_user.instruct.enabled && power_user.instruct.stop_sequence) {
-    return message_already_generated.indexOf(nameString) === -1 && //if there is no 'You:' in the response msg
+        if (message_already_generated.indexOf(power_user.instruct.stop_sequence) !== -1) {
-        message_already_generated.indexOf('<|endoftext|>') === -1 && //if there is no <endoftext> stamp in the response msg
+            return false;
-        tokens_already_generated < parseInt(amount_gen) && //if the gen'd msg is less than the max response length..
+        }
-        getMessage.length > 0;     //if we actually have gen'd text at all...
+    }
+
+    // stopping name string
+    const nameString = isImpersonate ? `${name2}:` : (is_pygmalion ? 'You:' : `${name1}:`);
+    // if there is no 'You:' in the response msg
+    const doesNotContainName = message_already_generated.indexOf(nameString) === -1;
+    //if there is no <endoftext> stamp in the response msg
+    const isNotEndOfText = message_already_generated.indexOf('<|endoftext|>') === -1;
+    //if the gen'd msg is less than the max response length..
+    const notReachedMax = tokens_already_generated < parseInt(amount_gen);
+    //if we actually have gen'd text at all...
+    const msgHasText = getMessage.length > 0;
+    return doesNotContainName && isNotEndOfText && notReachedMax && msgHasText;
 }
 
 function extractNameFromMessage(getMessage, force_name2, isImpersonate) {
@@ -2745,6 +2779,12 @@ function cleanUpMessage(getMessage, isImpersonate) {
             getMessage = getMessage.substring(0, getMessage.indexOf(power_user.instruct.stop_sequence));
         }
     }
+    if (power_user.instruct.enabled && power_user.instruct.input_sequence && isImpersonate) {
+        getMessage = getMessage.replaceAll(power_user.instruct.input_sequence, '');
+    }
+    if (power_user.instruct.enabled && power_user.instruct.output_sequence && !isImpersonate) {
+        getMessage = getMessage.replaceAll(power_user.instruct.output_sequence, '');
+    }
     // clean-up group message from excessive generations
     if (selected_group) {
         getMessage = cleanGroupMessage(getMessage);

public/scripts/extensions/stable-diffusion/index.js

@@ -237,9 +237,17 @@ function getQuietPrompt(mode, trigger) {
 function processReply(str) {
     str = str.replaceAll('"', '')
     str = str.replaceAll('“', '')
-    str = str.replaceAll('\n', ' ')
+    str = str.replaceAll('\n', ', ')
+    str = str.replace(/[^a-zA-Z0-9,:]+/g, ' ') // Replace everything except alphanumeric characters and commas with spaces
+    str = str.replace(/\s+/g, ' '); // Collapse multiple whitespaces into one
     str = str.trim();
 
+    str = str
+        .split(',') // list split by commas
+        .map(x => x.trim()) // trim each entry
+        .filter(x => x) // remove empty entries
+        .join(', '); // join it back with proper spacing
+
     return str;
 }
 
@@ -259,7 +267,7 @@ async function generatePicture(_, trigger) {
         const prompt = processReply(await new Promise(
             async function promptPromise(resolve, reject) {
                 try {
-                    await context.generate('quiet', { resolve, reject, quiet_prompt });
+                    await context.generate('quiet', { resolve, reject, quiet_prompt, force_name2: true, });
                 }
                 catch {
                     reject();
@@ -269,6 +277,8 @@ async function generatePicture(_, trigger) {
         context.deactivateSendButtons();
         hideSwipeButtons();
 
+        console.log('Processed Stable Diffusion prompt:', prompt);
+
         const url = new URL(getApiUrl());
         url.pathname = '/api/image';
         const result = await fetch(url, {
@@ -295,7 +305,7 @@ async function generatePicture(_, trigger) {
             sendMessage(prompt, base64Image);
         }
     } catch (err) {
-        console.error(err);
+        console.trace(err);
         throw new Error('SD prompt text generation failed.')
     }
     finally {