jaakko/agentic-studio
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Web Worker: WASM-inferenssi erillisessä säikeessä, UI ei jäädy

Browse Source 
- Poistettu kaikki web_sys::window() -kutsut Rust WASM:sta
- Uudet Worker-yhteensopivat apufunktiot: perf_now(), worker_fetch(), sleep_ms()
- worker.js lataa ja ajaa WASM-moduulin erillisessä säikeessä
- ensureCoderNode käynnistää Workerin pääsäikeen sijaan
- Selaimen UI pysyy responsiivisena inferenssin aikana

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
jaakko
2026-04-06 19:59:09 +03:00
parent fc95cf8c1b
commit b3646ae5d3
6 changed files with 129 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
network-poc/node/src/smollm.rs
View File

@@ -28,10 +28,7 @@ async fn ensure_cached(key: &str, url: &str, ws: &Rc<RefCell<WebSocket>>) -> Res
send_progress(ws, key, 0, 0, 0);
// Fetch API:lla saadaan Content-Length ja streaming-luku
let window = web_sys::window().unwrap();
let resp_val = wasm_bindgen_futures::JsFuture::from(window.fetch_with_str(url))
.await.map_err(|e| format!("Fetch epäonnistui: {:?}", e))?;
let resp: web_sys::Response = resp_val.dyn_into().map_err(|_| "Ei Response-objekti".to_string())?;
let resp = crate::worker_fetch(url).await?;
if !resp.ok() {
return Err(format!("HTTP {}", resp.status()));
@@ -99,7 +96,7 @@ fn send_progress(ws: &Rc<RefCell<WebSocket>>, file: &str, pct: u32, loaded: usiz
/// Lataa malli ja tokenizer, suorita inferenssi ja streamaa tokenit hubille
pub async fn run_smollm_inference(prompt: String, ws: Rc<RefCell<WebSocket>>) {
let perf = web_sys::window().unwrap().performance().unwrap();
// performance via crate::perf_now()
// 1. Lataa tokenizer
let tok_bytes = match ensure_cached("smollm-tokenizer.json", TOKENIZER_URL, &ws).await {
@@ -122,7 +119,7 @@ pub async fn run_smollm_inference(prompt: String, ws: Rc<RefCell<WebSocket>>) {
// Burn 0.21-pre.2 cubecl-runtime ei käänny Wasmille (println! puuttuu)
// → NdArray kunnes Burn 0.21 stable + Wasm-tuki
console_log!("[SmolLM] Burn NdArray (CPU) inferenssi...");
run_burn_inference::<burn::backend::NdArray>(prompt, model_bytes, tokenizer, ws, perf.clone()).await;
run_burn_inference::<burn::backend::NdArray>(prompt, model_bytes, tokenizer, ws).await;
}
async fn run_burn_inference<B: burn::tensor::backend::Backend>(
@@ -130,9 +127,8 @@ async fn run_burn_inference<B: burn::tensor::backend::Backend>(
model_bytes: Vec<u8>,
tokenizer: tokenizers::Tokenizer,
ws: Rc<RefCell<WebSocket>>,
perf: web_sys::Performance, // Korjattu Wasm-performanssi välitettäväksi
) {
let start_load = perf.now();
let start_load = crate::perf_now();
let device = Default::default();
let config = crate::burn_smollm::config::SmolLMConfig::default();
@@ -143,7 +139,7 @@ async fn run_burn_inference<B: burn::tensor::backend::Backend>(
Err(e) => { console_log!("[SmolLM] Lataus epäonnistui: {}", e); return; }
};
let load_time = perf.now() - start_load;
let load_time = crate::perf_now() - start_load;
console_log!("[SmolLM] Burn-malli ladattu ({:.0}ms). Generoidaan...", load_time);
let formatted_prompt = format!("<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n", prompt);
@@ -156,7 +152,7 @@ async fn run_burn_inference<B: burn::tensor::backend::Backend>(
let input_len = input_ids.len();
console_log!("[SmolLM] Syöte: {} tokenia", input_len);
let start_gen = perf.now();
let start_gen = crate::perf_now();
let max_new_tokens = 32;
let mut generated_text = String::new();
let mut tokens_generated: usize = 0;
@@ -219,7 +215,7 @@ async fn run_burn_inference<B: burn::tensor::backend::Backend>(
tokens_generated += 1;
}
let gen_time = perf.now() - start_gen;
let gen_time = crate::perf_now() - start_gen;
let tokens_per_sec = if gen_time > 0.0 { (tokens_generated as f64 / gen_time) * 1000.0 } else { 0.0 };
let done = serde_json::json!({