Spaces:
Running
Running
martia_f
commited on
Commit
·
b519309
1
Parent(s):
1bb6980
Download indicator + Removed yapping
Browse files- .DS_Store +0 -0
- index.html +1 -1
- phiWorker.js +167 -129
.DS_Store
ADDED
|
Binary file (6.15 kB). View file
|
|
|
index.html
CHANGED
|
@@ -184,7 +184,7 @@ Very polite review:`,
|
|
| 184 |
switch (data.status) {
|
| 185 |
case "loading":
|
| 186 |
outStatus.hidden = false;
|
| 187 |
-
outStatus.
|
| 188 |
outGen.hidden = true;
|
| 189 |
outCounter.hidden = true;
|
| 190 |
break;
|
|
|
|
| 184 |
switch (data.status) {
|
| 185 |
case "loading":
|
| 186 |
outStatus.hidden = false;
|
| 187 |
+
outStatus.innerHTML = data.message.replaceAll('\n', '<br>\n');
|
| 188 |
outGen.hidden = true;
|
| 189 |
outCounter.hidden = true;
|
| 190 |
break;
|
phiWorker.js
CHANGED
|
@@ -1,155 +1,193 @@
|
|
| 1 |
import init, { Model } from "./build/m.js";
|
| 2 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
async function fetchArrayBuffer(url) {
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
}
|
|
|
|
| 15 |
async function concatenateArrayBuffers(urls) {
|
| 16 |
-
|
| 17 |
|
| 18 |
-
|
| 19 |
-
|
| 20 |
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
}
|
| 28 |
|
| 29 |
class Phi {
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
quantized
|
| 38 |
-
) {
|
| 39 |
-
// load individual modelID only once
|
| 40 |
-
if (!this.instance[modelID]) {
|
| 41 |
-
await init();
|
| 42 |
-
|
| 43 |
-
self.postMessage({ status: "loading", message: "Loading Model" });
|
| 44 |
-
const [weightsArrayU8, tokenizerArrayU8, configArrayU8] =
|
| 45 |
-
await Promise.all([
|
| 46 |
-
weightsURL instanceof Array ? concatenateArrayBuffers(weightsURL) : fetchArrayBuffer(weightsURL),
|
| 47 |
-
fetchArrayBuffer(tokenizerURL),
|
| 48 |
-
fetchArrayBuffer(configURL),
|
| 49 |
-
]);
|
| 50 |
-
|
| 51 |
-
this.instance[modelID] = new Model(
|
| 52 |
-
weightsArrayU8,
|
| 53 |
-
tokenizerArrayU8,
|
| 54 |
-
configArrayU8,
|
| 55 |
quantized
|
| 56 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 57 |
}
|
| 58 |
-
return this.instance[modelID];
|
| 59 |
-
}
|
| 60 |
}
|
| 61 |
|
| 62 |
let controller = null;
|
| 63 |
self.addEventListener("message", (event) => {
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
});
|
| 71 |
|
| 72 |
async function generate(data) {
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 120 |
}
|
| 121 |
-
|
| 122 |
-
if (token === "<|endoftext|>") {
|
| 123 |
-
self.postMessage({
|
| 124 |
status: "complete",
|
| 125 |
message: "complete",
|
| 126 |
output: prompt + sentence,
|
| 127 |
-
});
|
| 128 |
-
return;
|
| 129 |
-
}
|
| 130 |
-
const tokensSec =
|
| 131 |
-
((tokensCount + 1) / (performance.now() - startTime)) * 1000;
|
| 132 |
-
|
| 133 |
-
sentence += token;
|
| 134 |
-
self.postMessage({
|
| 135 |
-
status: "generating",
|
| 136 |
-
message: "Generating token",
|
| 137 |
-
token: token,
|
| 138 |
-
sentence: sentence,
|
| 139 |
-
totalTime: performance.now() - startTime,
|
| 140 |
-
tokensSec,
|
| 141 |
-
prompt: prompt,
|
| 142 |
});
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
tokensCount++;
|
| 146 |
}
|
| 147 |
-
self.postMessage({
|
| 148 |
-
status: "complete",
|
| 149 |
-
message: "complete",
|
| 150 |
-
output: prompt + sentence,
|
| 151 |
-
});
|
| 152 |
-
} catch (e) {
|
| 153 |
-
self.postMessage({ error: e });
|
| 154 |
-
}
|
| 155 |
}
|
|
|
|
| 1 |
import init, { Model } from "./build/m.js";
|
| 2 |
|
| 3 |
+
function fixTwo(x) { return Math.floor(x * 100) / 100 }
|
| 4 |
+
|
| 5 |
+
function humanSize(size) {
|
| 6 |
+
if(size < 1e3) return `${fixTwo(size)}b`
|
| 7 |
+
if(size < 1e6) return `${fixTwo(size/1e3)}kb`
|
| 8 |
+
if(size < 1e9) return `${fixTwo(size/1e6)}mb`
|
| 9 |
+
if(size < 1e12) return `${fixTwo(size/1e9)}gb`
|
| 10 |
+
return `${fixTwo(size/1e12)}tb`
|
| 11 |
+
}
|
| 12 |
+
|
| 13 |
+
let lastSend = 0
|
| 14 |
+
|
| 15 |
async function fetchArrayBuffer(url) {
|
| 16 |
+
const cacheName = "phi-mixformer-candle-cache";
|
| 17 |
+
const cache = await caches.open(cacheName);
|
| 18 |
+
const cachedResponse = await cache.match(url);
|
| 19 |
+
if (cachedResponse) {
|
| 20 |
+
const data = await cachedResponse.arrayBuffer();
|
| 21 |
+
return new Uint8Array(data);
|
| 22 |
+
}
|
| 23 |
+
const res = await fetch(url, { cache: "force-cache" });
|
| 24 |
+
while (!res.body) { }
|
| 25 |
+
const reader = res.body.getReader();
|
| 26 |
+
const contentLength = +(res.headers.get('Content-Length') ?? 0);
|
| 27 |
+
let receivedLength = 0;
|
| 28 |
+
let chunks = [];
|
| 29 |
+
while (true) {
|
| 30 |
+
const { done, value } = await reader.read();
|
| 31 |
+
if (done) {
|
| 32 |
+
break;
|
| 33 |
+
}
|
| 34 |
+
chunks.push(value);
|
| 35 |
+
receivedLength += value.length;
|
| 36 |
+
let downloadMessage = `Downloading... ${fixTwo((receivedLength / contentLength) * 100)}% (${humanSize(Math.floor(receivedLength * 100) / 100)})\nLink: ${url}\nTotal size: ${humanSize(fixTwo(contentLength))}`
|
| 37 |
+
if(Date.now() - lastSend > 250) {
|
| 38 |
+
self.postMessage({ status: "loading", message: downloadMessage })
|
| 39 |
+
console.log(downloadMessage)
|
| 40 |
+
lastSend = Date.now()
|
| 41 |
+
}
|
| 42 |
+
}
|
| 43 |
+
let chunksAll = new Uint8Array(receivedLength);
|
| 44 |
+
let position = 0;
|
| 45 |
+
for (let chunk of chunks) {
|
| 46 |
+
chunksAll.set(chunk, position);
|
| 47 |
+
position += chunk.length;
|
| 48 |
+
}
|
| 49 |
+
cache.put(url, new Response(chunksAll));
|
| 50 |
+
return chunksAll;
|
| 51 |
}
|
| 52 |
+
|
| 53 |
async function concatenateArrayBuffers(urls) {
|
| 54 |
+
const arrayBuffers = await Promise.all(urls.map(url => fetchArrayBuffer(url)));
|
| 55 |
|
| 56 |
+
let totalLength = arrayBuffers.reduce((acc, arrayBuffer) => acc + arrayBuffer.byteLength, 0);
|
| 57 |
+
let concatenatedBuffer = new Uint8Array(totalLength);
|
| 58 |
|
| 59 |
+
let offset = 0;
|
| 60 |
+
arrayBuffers.forEach(buffer => {
|
| 61 |
+
concatenatedBuffer.set(new Uint8Array(buffer), offset);
|
| 62 |
+
offset += buffer.byteLength;
|
| 63 |
+
});
|
| 64 |
+
return concatenatedBuffer;
|
| 65 |
}
|
| 66 |
|
| 67 |
class Phi {
|
| 68 |
+
static instance = {};
|
| 69 |
+
|
| 70 |
+
static async getInstance(
|
| 71 |
+
weightsURL,
|
| 72 |
+
modelID,
|
| 73 |
+
tokenizerURL,
|
| 74 |
+
configURL,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 75 |
quantized
|
| 76 |
+
) {
|
| 77 |
+
// load individual modelID only once
|
| 78 |
+
if (!this.instance[modelID]) {
|
| 79 |
+
await init();
|
| 80 |
+
|
| 81 |
+
self.postMessage({ status: "loading", message: "Loading Model" });
|
| 82 |
+
const [weightsArrayU8, tokenizerArrayU8, configArrayU8] =
|
| 83 |
+
await Promise.all([
|
| 84 |
+
weightsURL instanceof Array ? concatenateArrayBuffers(weightsURL) : fetchArrayBuffer(weightsURL),
|
| 85 |
+
fetchArrayBuffer(tokenizerURL),
|
| 86 |
+
fetchArrayBuffer(configURL),
|
| 87 |
+
]);
|
| 88 |
+
|
| 89 |
+
this.instance[modelID] = new Model(
|
| 90 |
+
weightsArrayU8,
|
| 91 |
+
tokenizerArrayU8,
|
| 92 |
+
configArrayU8,
|
| 93 |
+
quantized
|
| 94 |
+
);
|
| 95 |
+
}
|
| 96 |
+
return this.instance[modelID];
|
| 97 |
}
|
|
|
|
|
|
|
| 98 |
}
|
| 99 |
|
| 100 |
let controller = null;
|
| 101 |
self.addEventListener("message", (event) => {
|
| 102 |
+
if (event.data.command === "start") {
|
| 103 |
+
controller = new AbortController();
|
| 104 |
+
generate(event.data);
|
| 105 |
+
} else if (event.data.command === "abort") {
|
| 106 |
+
controller.abort();
|
| 107 |
+
}
|
| 108 |
});
|
| 109 |
|
| 110 |
async function generate(data) {
|
| 111 |
+
const {
|
| 112 |
+
weightsURL,
|
| 113 |
+
modelID,
|
| 114 |
+
tokenizerURL,
|
| 115 |
+
configURL,
|
| 116 |
+
quantized,
|
| 117 |
+
prompt,
|
| 118 |
+
temp,
|
| 119 |
+
top_p,
|
| 120 |
+
repeatPenalty,
|
| 121 |
+
seed,
|
| 122 |
+
maxSeqLen,
|
| 123 |
+
} = data;
|
| 124 |
+
try {
|
| 125 |
+
self.postMessage({ status: "loading", message: "Starting Phi" });
|
| 126 |
+
const model = await Phi.getInstance(
|
| 127 |
+
weightsURL,
|
| 128 |
+
modelID,
|
| 129 |
+
tokenizerURL,
|
| 130 |
+
configURL,
|
| 131 |
+
quantized
|
| 132 |
+
);
|
| 133 |
+
|
| 134 |
+
self.postMessage({ status: "loading", message: "Initializing model" });
|
| 135 |
+
const firstToken = model.init_with_prompt(
|
| 136 |
+
prompt,
|
| 137 |
+
temp,
|
| 138 |
+
top_p,
|
| 139 |
+
repeatPenalty,
|
| 140 |
+
64,
|
| 141 |
+
BigInt(seed)
|
| 142 |
+
);
|
| 143 |
+
const seq_len = 2048;
|
| 144 |
+
|
| 145 |
+
let sentence = firstToken;
|
| 146 |
+
let maxTokens = maxSeqLen ? maxSeqLen : seq_len - prompt.length - 1;
|
| 147 |
+
let startTime = performance.now();
|
| 148 |
+
let tokensCount = 0;
|
| 149 |
+
while (tokensCount < maxTokens) {
|
| 150 |
+
await new Promise(async (resolve) => {
|
| 151 |
+
if (controller && controller.signal.aborted) {
|
| 152 |
+
self.postMessage({
|
| 153 |
+
status: "aborted",
|
| 154 |
+
message: "Aborted",
|
| 155 |
+
output: prompt + sentence,
|
| 156 |
+
});
|
| 157 |
+
return;
|
| 158 |
+
}
|
| 159 |
+
const token = await model.next_token();
|
| 160 |
+
if (token === "<|endoftext|>") {
|
| 161 |
+
self.postMessage({
|
| 162 |
+
status: "complete",
|
| 163 |
+
message: "complete",
|
| 164 |
+
output: prompt + sentence,
|
| 165 |
+
});
|
| 166 |
+
return;
|
| 167 |
+
}
|
| 168 |
+
const tokensSec =
|
| 169 |
+
((tokensCount + 1) / (performance.now() - startTime)) * 1000;
|
| 170 |
+
|
| 171 |
+
sentence += token;
|
| 172 |
+
self.postMessage({
|
| 173 |
+
status: "generating",
|
| 174 |
+
message: "Generating token",
|
| 175 |
+
token: token,
|
| 176 |
+
sentence: sentence,
|
| 177 |
+
totalTime: performance.now() - startTime,
|
| 178 |
+
tokensSec,
|
| 179 |
+
prompt: prompt,
|
| 180 |
+
});
|
| 181 |
+
setTimeout(resolve, 0);
|
| 182 |
+
});
|
| 183 |
+
tokensCount++;
|
| 184 |
}
|
| 185 |
+
self.postMessage({
|
|
|
|
|
|
|
| 186 |
status: "complete",
|
| 187 |
message: "complete",
|
| 188 |
output: prompt + sentence,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 189 |
});
|
| 190 |
+
} catch (e) {
|
| 191 |
+
self.postMessage({ error: e });
|
|
|
|
| 192 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 193 |
}
|