Compare commits

...
Sign in to create a new pull request.

12 commits

Author SHA1 Message Date
Henri Vasserman
28046d1e52
Merge and update 2023-08-09 00:36:11 +03:00
Henri Vasserman
ca2467d12c
chat css 2023-07-24 14:09:05 +03:00
Henri Vasserman
f77972f9af
Merge remote-tracking branch 'origin/master' into server-cfg 2023-07-24 14:08:54 +03:00
Henri Vasserman
e4db70720d
[wip] chat now has parameter and cfg 2023-07-20 15:37:31 +03:00
Henri Vasserman
082dd81286
[wip] chat improvements 2023-07-20 03:48:48 +03:00
Henri Vasserman
43694ca867
consistent semicolons 2023-07-20 00:58:16 +03:00
Henri Vasserman
890d1b8446
Merge master into server-cfg 2023-07-20 00:48:03 +03:00
Henri Vasserman
dd3cf5760a
last n tokens done 2023-07-20 00:36:36 +03:00
Henri Vasserman
42591a0acd
remove "smooth factor" 2023-07-20 00:02:13 +03:00
Henri Vasserman
2cb8469e7f
refactor evaluation logic 2023-07-19 23:45:40 +03:00
Henri Vasserman
4cae9f5673
Port CFG to server. 2023-07-13 22:37:57 +03:00
Henri Vasserman
3a13d1e829
Apply formatting. 2023-07-13 20:24:31 +03:00
3 changed files with 2087 additions and 1788 deletions

File diff suppressed because it is too large Load diff

View file

@ -10,6 +10,11 @@
body { body {
font-family: system-ui; font-family: system-ui;
font-size: 90%; font-size: 90%;
max-width: 600px;
min-width: 300px;
line-height: 1.2;
margin: 0 auto;
padding: 0 0.5em;
} }
#container { #container {
@ -35,12 +40,37 @@
padding: 0.5em; padding: 0.5em;
} }
body { #chat {
max-width: 600px; display: flex;
min-width: 300px; flex-direction: column;
line-height: 1.2; }
margin: 0 auto;
padding: 0 0.5em; .bubble {
border: 1px solid;
border-radius: 1.0em;
padding: 0.5em;
max-width: 75%;
margin-top: 0;
}
.user {
background-color: #161616;
color: #d6d6d6;
border-bottom-right-radius: 0;
place-self: flex-end;
}
.asst {
background-color: #d6d6d6;
color: #161616;
border-bottom-left-radius: 0;
place-self: flex-start;
}
.typing {
color: #888;
text-align: left;
font-size: 120%;
} }
p { p {
@ -102,7 +132,6 @@
padding: 0.5em; padding: 0.5em;
} }
textarea { textarea {
padding: 5px; padding: 5px;
flex-grow: 1; flex-grow: 1;
@ -138,34 +167,35 @@
<script type="module"> <script type="module">
import { import {
html, h, signal, effect, computed, render, useSignal, useEffect, useRef html, h, signal, effect, computed, render, useSignal, useEffect, useRef
} from '/index.js'; } from '/index.js'
import { llama } from '/completion.js'; import { llama } from '/completion.js'
const session = signal({ const session = signal({
prompt: "This is a conversation between user and llama, a friendly chatbot. respond in simple markdown.", system: "A chat between a curious user and a pirate.",
template: "{{prompt}}\n\n{{history}}\n{{char}}:", system_cfg: "A chat between a curious user and an artificial intelligence assistant.",
historyTemplate: "{{name}}: {{message}}", message: "{{system}}\n\n### Instruction:\n{{user}}\n\n### Response:\n{{assistant}}",
stop: ["###"],
transcript: [], transcript: [],
type: "chat", type: "chat",
char: "llama",
user: "User",
}) })
const params = signal({ const params = signal({
cfg_scale: 4.0,
frequency_penalty: 0.0, // 0.0 = disabled
mirostat_eta: 0.1, // learning rate
mirostat_tau: 5, // target entropy
mirostat: 0, // 0/1/2
n_predict: 400, n_predict: 400,
temperature: 0.7, penalize_nl: true,
presence_penalty: 0.0, // 0.0 = disabled
repeat_last_n: 256, // 0 = disable penalty, -1 = context size repeat_last_n: 256, // 0 = disable penalty, -1 = context size
repeat_penalty: 1.18, // 1.0 = disabled repeat_penalty: 1.18, // 1.0 = disabled
temperature: 0.7,
tfs_z: 1.0, // 1.0 = disabled
top_k: 40, // <= 0 to use vocab size top_k: 40, // <= 0 to use vocab size
top_p: 0.5, // 1.0 = disabled top_p: 0.5, // 1.0 = disabled
tfs_z: 1.0, // 1.0 = disabled
typical_p: 1.0, // 1.0 = disabled typical_p: 1.0, // 1.0 = disabled
presence_penalty: 0.0, // 0.0 = disabled
frequency_penalty: 0.0, // 0.0 = disabled
mirostat: 0, // 0/1/2
mirostat_tau: 5, // target entropy
mirostat_eta: 0.1, // learning rate
}) })
const llamaStats = signal(null) const llamaStats = signal(null)
@ -177,88 +207,92 @@
const transcriptUpdate = (transcript) => { const transcriptUpdate = (transcript) => {
session.value = { session.value = {
...session.value, ...session.value,
transcript transcript,
} }
} }
// simple template replace // simple template replace
const template = (str, extraSettings) => { const template = (str, extraSettings) => {
let settings = session.value; let settings = session.value
if (extraSettings) { if (extraSettings) {
settings = { ...settings, ...extraSettings }; settings = { ...settings, ...extraSettings }
} }
return String(str).replaceAll(/\{\{(.*?)\}\}/g, (_, key) => template(settings[key])); return String(str).replaceAll(/\{\{(.*?)\}\}/g, (_, key) => template(settings[key]))
} }
// send message to server // send message to server
const chat = async (msg) => { const chat = async (msg) => {
if (controller.value) { if (controller.value) {
console.log('already running...'); console.log('already running...')
return; return
} }
controller.value = new AbortController(); controller.value = new AbortController()
transcriptUpdate([...session.value.transcript, ["{{user}}", msg]])
const prompt = template(session.value.template, {
message: msg,
history: session.value.transcript.flatMap(([name, message]) => template(session.value.historyTemplate, {name, message})).join("\n"),
});
let currentMessage = '';
const history = session.value.transcript const history = session.value.transcript
const system = history.length == 0 ? session.value.system : ""
transcriptUpdate([...history, { system, user: msg, assistant: "" }])
const prompt = session.value.transcript.map(t =>
template(session.value.message, t)).join("").trimEnd()
const cfg_negative_prompt = params.value.cfg_scale > 1 ? session.value.transcript.map(t =>
template(session.value.message, { ...t, system: session.value.system_cfg })
).join("").trimEnd() : ""
let currentMessage = ''
const llamaParams = { const llamaParams = {
...params.value, ...params.value,
stop: ["</s>", template("{{char}}:"), template("{{user}}:")], cfg_negative_prompt,
stop: session.stop,
} }
for await (const chunk of llama(prompt, llamaParams, { controller: controller.value })) { for await (const chunk of llama(prompt, llamaParams, { controller: controller.value })) {
const data = chunk.data; const data = chunk.data
currentMessage += data.content; currentMessage += data.content
// remove leading whitespace // remove leading whitespace
currentMessage = currentMessage.replace(/^\s+/, "") currentMessage = currentMessage.replace(/^\s+/, "")
transcriptUpdate([...history, ["{{char}}", currentMessage]]) transcriptUpdate([...history, { system, user: msg, assistant: currentMessage }])
if (data.stop) { if (data.stop) {
console.log("Completion finished: '", currentMessage, "', summary: ", data); console.log("Completion finished: '", currentMessage, "', summary: ", data)
} }
if (data.timings) { if (data.timings) {
llamaStats.value = data.timings; llamaStats.value = data.timings
} }
} }
controller.value = null; controller.value = null
} }
function MessageInput() { function MessageInput() {
const message = useSignal("") const message = useSignal("")
const stop = (e) => { const stop = (e) => {
e.preventDefault(); e.preventDefault()
if (controller.value) { if (controller.value) {
controller.value.abort(); controller.value.abort()
controller.value = null; controller.value = null
} }
} }
const reset = (e) => { const reset = (e) => {
stop(e); stop(e)
transcriptUpdate([]); transcriptUpdate([])
} }
const submit = (e) => { const submit = (e) => {
stop(e); stop(e)
chat(message.value); chat(message.value)
message.value = ""; message.value = ""
} }
const enterSubmits = (event) => { const enterSubmits = (event) => {
if (event.which === 13 && !event.shiftKey) { if (event.which === 13 && !event.shiftKey) {
submit(event); submit(event)
} }
} }
@ -277,7 +311,7 @@
} }
const ChatLog = (props) => { const ChatLog = (props) => {
const messages = session.value.transcript; const messages = session.value.transcript
const container = useRef(null) const container = useRef(null)
useEffect(() => { useEffect(() => {
@ -287,102 +321,132 @@
} }
}, [messages]) }, [messages])
const chatLine = ([user, msg]) => {
return html`<p key=${msg}><strong>${template(user)}:</strong> <${Markdownish} text=${template(msg)} /></p>`
};
return html` return html`
<section id="chat" ref=${container}> <section id="chat" ref=${container}>
${messages.flatMap(chatLine)} ${messages.map(({system, user, assistant}) => html`
</section>`; ${system !== "" && html`<p><em><${Markdownish} text=${system} /></em></p>`}
<p class="user bubble"><${Markdownish} text=${user} /></p>
${assistant !== "" ?
html`<p class="asst bubble"><${Markdownish} text=${assistant} /></p>` :
html`<p class="typing">...</p>`}
`)}
</section>
`
}
const updateSession = (el) => session.value = { ...session.value, [el.target.name]: el.target.value }
const updateParams = (el) => params.value = { ...params.value, [el.target.name]: el.target.value }
const updateParamsFloat = (el) => params.value = { ...params.value, [el.target.name]: parseFloat(el.target.value) }
const updateParamsInt = (el) => params.value = { ...params.value, [el.target.name]: Math.floor(parseFloat(el.target.value)) }
const updateArray = (el) => {
const [name, index] = el.target.name.split(".")
const newarr = session.value[name].map((v, i) => i == index ? el.target.value : v).filter(x => x !== "")
session.value = { ...session.value, [name]: newarr }
}
const appendArray = () => session.value = { ...session.value, stop: [...session.value.stop, ""] }
const ParamSlider = ({param, min, max, step, children}) => {
return html`
<div>
<label for="${param}">${children}</label>
<input type="range" id="${param}" min="${min}" max="${max}" step="${step}" name="${param}" value="${params.value[param]}" oninput=${updateParamsFloat} />
<span>${params.value[param]}</span>
</div>
`
}
const FloatField = ({label, max, min, name, step, value}) => {
return html`
<div>
<label for="${name}">${label}</label>
<input type="range" id="${name}" min="${min}" max="${max}" step="${step}" name="${name}" value="${value}" oninput=${updateParamsFloat} />
<span>${value}</span>
</div>
`
};
const IntField = ({label, max, min, name, value}) => {
return html`
<div>
<label for="${name}">${label}</label>
<input type="range" id="${name}" min="${min}" max="${max}" name="${name}" value="${value}" oninput=${updateParamsInt} />
<span>${value}</span>
</div>
`
}; };
const ConfigForm = (props) => { const ConfigForm = (props) => {
const updateSession = (el) => session.value = { ...session.value, [el.target.name]: el.target.value }
const updateParams = (el) => params.value = { ...params.value, [el.target.name]: el.target.value }
const updateParamsFloat = (el) => params.value = { ...params.value, [el.target.name]: parseFloat(el.target.value) }
const updateParamsInt = (el) => params.value = { ...params.value, [el.target.name]: Math.floor(parseFloat(el.target.value)) }
const FloatField = ({label, max, min, name, step, value}) => {
return html`
<div>
<label for="${name}">${label}</label>
<input type="range" id="${name}" min="${min}" max="${max}" step="${step}" name="${name}" value="${value}" oninput=${updateParamsFloat} />
<span>${value}</span>
</div>
`
};
const IntField = ({label, max, min, name, value}) => {
return html`
<div>
<label for="${name}">${label}</label>
<input type="range" id="${name}" min="${min}" max="${max}" name="${name}" value="${value}" oninput=${updateParamsInt} />
<span>${value}</span>
</div>
`
};
return html` return html`
<form> <form>
<fieldset> <fieldset>
<div> <div>
<label for="prompt">Prompt</label> <label for="system">System prompt</label>
<textarea type="text" name="prompt" value="${session.value.prompt}" rows=4 oninput=${updateSession}/> <textarea type="text" name="system" value="${session.value.system}" rows=4 oninput=${updateSession}/>
</div>
</fieldset>
<fieldset class="two">
<div>
<label for="user">User name</label>
<input type="text" name="user" value="${session.value.user}" oninput=${updateSession} />
</div> </div>
<div> <div>
<label for="bot">Bot name</label> <label for="message">Message template</label>
<input type="text" name="char" value="${session.value.char}" oninput=${updateSession} /> <textarea type="text" name="message" value="${session.value.message}" rows=7 oninput=${updateSession}/>
</div>
</fieldset>
<fieldset>
<div>
<label for="template">Prompt template</label>
<textarea id="template" name="template" value="${session.value.template}" rows=4 oninput=${updateSession}/>
</div> </div>
<div> <div>
<label for="template">Chat history template</label> <label for="stop">Stop strings</label>
<textarea id="template" name="historyTemplate" value="${session.value.historyTemplate}" rows=1 oninput=${updateSession}/> ${session.value.stop.map((stop, i) => html`
<p><input type="text" name="stop.${i}" value="${stop}" oninput=${updateArray}/></p>
`)}
<input type="button" value="+" onclick=${appendArray} />
</div> </div>
</fieldset>
<fieldset class="two"> <${ParamSlider} min=1 max=10 step=0.1 param=cfg_scale>CFG scale<//>
${IntField({label: "Predictions", max: 2048, min: -1, name: "n_predict", value: params.value.n_predict})} ${params.value.cfg_scale > 1 && html`
${FloatField({label: "Temperature", max: 1.5, min: 0.0, name: "temperature", step: 0.01, value: params.value.temperature})}
${FloatField({label: "Penalize repeat sequence", max: 2.0, min: 0.0, name: "repeat_penalty", step: 0.01, value: params.value.repeat_penalty})}
${IntField({label: "Consider N tokens for penalize", max: 2048, min: 0, name: "repeat_last_n", value: params.value.repeat_last_n})}
${IntField({label: "Top-K sampling", max: 100, min: -1, name: "top_k", value: params.value.top_k})}
${FloatField({label: "Top-P sampling", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p})}
</fieldset>
<details>
<summary>More options</summary>
<fieldset class="two">
${FloatField({label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z})}
${FloatField({label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p})}
${FloatField({label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty})}
${FloatField({label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty})}
</fieldset>
<hr />
<fieldset class="three">
<div> <div>
<label><input type="radio" name="mirostat" value="0" checked=${params.value.mirostat == 0} oninput=${updateParamsInt} /> no Mirostat</label> <label for="system_cfg">CFG System prompt</label>
<label><input type="radio" name="mirostat" value="1" checked=${params.value.mirostat == 1} oninput=${updateParamsInt} /> Mirostat v1</label> <textarea type="text" name="system_cfg" value="${session.value.system_cfg}" rows=4 oninput=${updateSession}/>
<label><input type="radio" name="mirostat" value="2" checked=${params.value.mirostat == 2} oninput=${updateParamsInt} /> Mirostat v2</label>
</div> </div>
${FloatField({label: "Mirostat tau", max: 10.0, min: 0.0, name: "mirostat_tau", step: 0.01, value: params.value.mirostat_tau})} `}
${FloatField({label: "Mirostat eta", max: 1.0, min: 0.0, name: "mirostat_eta", step: 0.01, value: params.value.mirostat_eta})} </fieldset>
<fieldset class="two">
<${ParamSlider} min=1 max=1000 step=1 param=n_predict>Predict N tokens<//>
<${ParamSlider} min=0 max=1000 step=1 param=repeat_last_n>Penalize last N tokens<//>
${params.value.repeat_last_n > 0 && html`
<${ParamSlider} min=0 max=4 step=0.01 param=repeat_penalty>Penalize repeat sequence<//>
<${ParamSlider} min=0 max=4 step=0.01 param=frequency_penalty>Penalize frequent tokens<//>
<${ParamSlider} min=0 max=4 step=0.01 param=presence_penalty>Penalize tokens not present in prompt<//>
`}
<${ParamSlider} min=0 max=2 step=0.01 param=temperature>Temperature<//>
</fieldset>
${params.value.temperature > 0 && html`
<fieldset class="three">
<label>
<input type=radio name=mirostat checked=${params.value.mirostat == 0} value=0 oninput=${updateParamsFloat} />
Temperature
</label>
<label><input type=radio name=mirostat checked=${params.value.mirostat == 1} value=1 oninput=${updateParamsFloat} />
Mirostat v1
</label>
<label>
<input type=radio name=mirostat checked=${params.value.mirostat == 2} value=2 oninput=${updateParamsFloat} />
Mirostat v2
</label>
</fieldset> </fieldset>
</details>
<fieldset class="two">
${params.value.mirostat == 0 && html`
<${ParamSlider} min=1 max=1000 step=1 param=top_k>Top K<//>
<${ParamSlider} min=0 max=1 step=0.01 param=tfs_z>Tail free sampling<//>
<${ParamSlider} min=0 max=1 step=0.01 param=typical_p>Typical P<//>
<${ParamSlider} min=0 max=1 step=0.01 param=top_p>Top P<//>
`}
${params.value.mirostat > 0 && html`
<${ParamSlider} min=0 max=1 step=0.01 param=mirostat_eta>Mirostat eta, learning rate<//>
<${ParamSlider} min=0 max=1000 step=1 param=mirostat_tau>Mirostat tau, target entropy<//>
`}
</fieldset>
`}
</form> </form>
` `
} }
@ -399,9 +463,9 @@
.replace(/_(.*?)_/g, '<em>$1</em>') .replace(/_(.*?)_/g, '<em>$1</em>')
.replace(/```.*?\n([\s\S]*?)```/g, '<pre><code>$1</code></pre>') .replace(/```.*?\n([\s\S]*?)```/g, '<pre><code>$1</code></pre>')
.replace(/`(.*?)`/g, '<code>$1</code>') .replace(/`(.*?)`/g, '<code>$1</code>')
.replace(/\n/gim, '<br />'); .replace(/\n/gim, '<br />')
return html`<span dangerouslySetInnerHTML=${{ __html: md }} />`; return html`<span dangerouslySetInnerHTML=${{ __html: md }} />`
}; }
const ModelGenerationInfo = (params) => { const ModelGenerationInfo = (params) => {
if (!llamaStats.value) { if (!llamaStats.value) {
@ -435,10 +499,10 @@
<p>Powered by <a href="https://github.com/ggerganov/llama.cpp">llama.cpp</a> and <a href="https://ggml.ai">ggml.ai</a>.</p> <p>Powered by <a href="https://github.com/ggerganov/llama.cpp">llama.cpp</a> and <a href="https://ggml.ai">ggml.ai</a>.</p>
</footer> </footer>
</div> </div>
`; `
} }
render(h(App), document.body); render(h(App), document.body)
</script> </script>
</head> </head>

File diff suppressed because it is too large Load diff