Versio 0.2.4

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

.gitignore

@@ -34,3 +34,6 @@ Cargo.lock
 *.pdb
 
 # End of https://www.toptal.com/developers/gitignore/api/rust,linux
+
+# Ajonaikaiset tietokannat
+*.db

network-poc/Dockerfile.native-node

@@ -1,30 +1,36 @@
 FROM rust:slim AS builder
 
 RUN apt-get update && apt-get install -y \
-    pkg-config libssl-dev g++ \
+    pkg-config libssl-dev g++ libvulkan-dev \
     && rm -rf /var/lib/apt/lists/*
 
 WORKDIR /app
-COPY Cargo.toml Cargo.lock ./
+COPY Cargo.toml ./
+COPY Cargo.loc[k] ./
 COPY hub/Cargo.toml hub/Cargo.toml
 COPY node/Cargo.toml node/Cargo.toml
 COPY native-node/Cargo.toml native-node/Cargo.toml
+COPY cli/Cargo.toml cli/Cargo.toml
 
 # Tyhjät src-tiedostot riippuvuuksien esikääntämistä varten
-RUN mkdir -p hub/src node/src native-node/src \
+RUN mkdir -p hub/src node/src native-node/src cli/src \
     && echo "fn main(){}" > hub/src/main.rs \
     && echo "" > node/src/lib.rs \
     && echo "fn main(){}" > native-node/src/main.rs \
+    && echo "fn main(){}" > cli/src/main.rs \
     && cargo build --release -p native-node 2>/dev/null || true
 
 COPY native-node/src native-node/src
-RUN cargo build --release -p native-node
+# Touch pakottaa rekompilauksen dummy-binaryn yli
+RUN touch native-node/src/main.rs && cargo build --release -p native-node
 
 FROM debian:bookworm-slim
-RUN apt-get update && apt-get install -y ca-certificates && rm -rf /var/lib/apt/lists/*
+RUN apt-get update && apt-get install -y ca-certificates libvulkan1 && rm -rf /var/lib/apt/lists/*
 COPY --from=builder /app/target/release/native-node /usr/local/bin/native-node
 
-ENV HUB_URL=ws://hub:3000/ws
+ENV HUB_URL=ws://agentic-poc:3000/ws
+ENV OLLAMA_URL=http://ollama:11434
+ENV OLLAMA_MODEL=qwen2.5-coder:7b
 ENV ALLOCATED_GB=4
 
 CMD ["native-node"]

network-poc/TODO.md

@@ -0,0 +1,26 @@
+# TODO — Kipinä Agentic Network
+
+## Turvallisuus
+- [ ] **Tulosten validointi** — solmu voi palauttaa haitallista koodia. Tarvitaan proof-of-work tai challenge-response -mekanismi
+- [ ] **Reputaatiojärjestelmä** — solmujen luotettavuuden seuranta: onnistuneet tehtävät, vasteaika, laatu
+- [ ] **Koodin sandboxaus** — generoitu koodi pitää ajaa eristetyssä ympäristössä ennen käyttäjälle näyttämistä
+- [ ] **Solmun identiteetti** — rekisteröityminen ja tunnistautuminen (API-avain / token)
+
+## Yksityisyys
+- [ ] **Promptien salaus** — käyttäjän promptit menevät tuntemattomalle solmulle selkotekstinä
+- [ ] **End-to-end enkryptio** — hub ei näe promptin sisältöä, vain reitittää
+- [ ] **Tietosuojaseloste** — käyttäjille kerrottava miten data kulkee ja kuka sen näkee
+- [ ] **Opt-in malli** — käyttäjä valitsee haluaako käyttää yhteisösolmuja vai vain omaa
+
+## Väärinkäytön esto
+- [ ] **Rate limiting per käyttäjä** — nykyinen IP-pohjainen ei riitä, tarvitaan autentikointi
+- [ ] **Solmun kuormitusraja** — solmu voi asettaa max tehtävät/minuutti
+- [ ] **Token-talous** — laskentaresurssien käyttö vaatii Kipinä-tokeneita (gamification jo aloitettu)
+- [ ] **Abuse reporting** — mekanismi haitallisten solmujen ilmiantamiseen
+
+## Seuraavat ominaisuudet
+- [ ] Agenttien välinen keskustelu (manageri ohjaa dynaamisesti)
+- [ ] Tehtävähistoria ja tulosten tallennus
+- [ ] Prometheus/OpenTelemetry -metriikat
+- [ ] Solmujen terveystarkistukset (ping/pong)
+- [ ] Streaming-vastaukset Ollaman kautta

network-poc/docker-compose.client.yml

@@ -1,13 +1,12 @@
 services:
-  # NVIDIA GPU -solmu
+  # Ollama NVIDIA GPU:lla
-  native-node-nvidia:
+  ollama-nvidia:
-    build:
+    image: ollama/ollama:latest
-      context: .
+    container_name: kipina-ollama
-      dockerfile: Dockerfile.native-node
+    ports:
-    container_name: kipina-node-nvidia
+      - "11434:11434"
-    environment:
+    volumes:
-      - HUB_URL=wss://kipina.studio/ws
+      - ollama-models:/root/.ollama
-      - ALLOCATED_GB=4
     restart: unless-stopped
     deploy:
       resources:
@@ -16,6 +15,65 @@ services:
             - driver: nvidia
               count: all
               capabilities: [gpu]
+    networks:
+      default:
+        aliases:
+          - ollama
+    profiles:
+      - nvidia
+
+  # Ollama AMD ROCm GPU:lla
+  ollama-amd:
+    image: ollama/ollama:rocm
+    container_name: kipina-ollama
+    ports:
+      - "11434:11434"
+    volumes:
+      - ollama-models:/root/.ollama
+    restart: unless-stopped
+    devices:
+      - /dev/kfd:/dev/kfd
+      - /dev/dri:/dev/dri
+    group_add:
+      - video
+      - render
+    networks:
+      default:
+        aliases:
+          - ollama
+    profiles:
+      - amd
+
+  # Ollama CPU:lla
+  ollama-cpu:
+    image: ollama/ollama:latest
+    container_name: kipina-ollama
+    ports:
+      - "11434:11434"
+    volumes:
+      - ollama-models:/root/.ollama
+    restart: unless-stopped
+    networks:
+      default:
+        aliases:
+          - ollama
+    profiles:
+      - cpu
+
+  # NVIDIA GPU -solmu
+  native-node-nvidia:
+    build:
+      context: .
+      dockerfile: Dockerfile.native-node
+    container_name: kipina-node-nvidia
+    environment:
+      - HUB_URL=wss://kipina.studio/ws
+      - OLLAMA_URL=http://ollama:11434
+      - OLLAMA_MODEL=qwen2.5-coder:7b
+      - ALLOCATED_GB=4
+    restart: unless-stopped
+    depends_on:
+      - ollama-nvidia
     profiles:
       - nvidia
 
@@ -27,14 +85,12 @@ services:
     container_name: kipina-node-amd
     environment:
       - HUB_URL=wss://kipina.studio/ws
+      - OLLAMA_URL=http://ollama:11434
+      - OLLAMA_MODEL=qwen2.5-coder:7b
       - ALLOCATED_GB=4
     restart: unless-stopped
-    devices:
+    depends_on:
-      - /dev/kfd:/dev/kfd
+      - ollama-amd
-      - /dev/dri:/dev/dri
-    group_add:
-      - video
-      - render
     profiles:
       - amd
 
@@ -46,7 +102,14 @@ services:
     container_name: kipina-node-cpu
     environment:
       - HUB_URL=wss://kipina.studio/ws
+      - OLLAMA_URL=http://ollama:11434
+      - OLLAMA_MODEL=qwen2.5-coder:7b
       - ALLOCATED_GB=2
     restart: unless-stopped
+    depends_on:
+      - ollama-cpu
     profiles:
       - cpu
+
+volumes:
+  ollama-models:

network-poc/docker-compose.yml

@@ -11,24 +11,37 @@ services:
     # Käännetään aina käynnistyksen yhteydessä varmuuden vuoksi Wasm uusimmista koodeista, ja päälle pyöräytetään Hub!
     command: bash -c "cd node && wasm-pack build --release --target web --out-dir ../static/pkg && cd ../hub && cargo run"
 
-  # Valinnainen natiivi-solmu — kerää oikeat laitteistotiedot (nvidia-smi-taso)
+  # Ollama — LLM-inferenssi
+  # NVIDIA: vaihda image → ollama/ollama:latest ja lisää deploy.resources (ks. README)
+  # CPU: vaihda image → ollama/ollama:latest ja poista devices
+  ollama:
+    image: ollama/ollama:rocm
+    container_name: kipina_ollama
+    ports:
+      - "11434:11434"
+    volumes:
+      - ollama-models:/root/.ollama
+    devices:
+      - /dev/kfd
+      - /dev/dri
+    profiles:
+      - native
+
+  # Natiivisolmu — yhdistää hubiin ja käyttää Ollamaa inferenssiin
   native-node:
     build:
       context: .
       dockerfile: Dockerfile.native-node
     container_name: kipina_native_node
     environment:
-      - HUB_URL=ws://agentic-poc:3000/ws
+      - HUB_URL=wss://kipina.studio/ws
+      - OLLAMA_URL=http://ollama:11434
+      - OLLAMA_MODEL=qwen2.5-coder:7b
       - ALLOCATED_GB=4
     depends_on:
-      - agentic-poc
+      - ollama
-    # GPU passthrough (valinnainen — toimii myös ilman)
-    deploy:
-      resources:
-        reservations:
-          devices:
-            - driver: nvidia
-              count: all
-              capabilities: [gpu]
     profiles:
       - native
+
+volumes:
+  ollama-models:

network-poc/hub/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "hub"
-version = "0.2.0"
+version = "0.2.4"
 edition = "2024"
 
 [dependencies]
@@ -16,3 +16,4 @@ futures = "0.3"
 rusqlite = { version = "0.31", features = ["bundled"] }
 chrono = "0.4"
 base64 = "0.22"
+reqwest = { version = "0.12", features = ["json"] }

network-poc/hub/src/main.rs

@@ -39,6 +39,7 @@ struct AppState {
     ip_connections: Mutex<HashMap<IpAddr, u32>>,
     node_ips: Mutex<HashMap<u64, IpAddr>>,
     node_tasks: Mutex<HashMap<u64, String>>,  // node_id → selected_task
+    node_types: Mutex<HashMap<u64, String>>,  // node_id → "native" | "browser"
     node_busy: Mutex<std::collections::HashSet<u64>>,  // Solmut joilla on aktiivinen tehtävä
     pending_task_ids: Mutex<std::collections::HashSet<String>>,  // Hubin jakamat task_id:t (gamification-validointi)
     api_rate_limits: Mutex<HashMap<IpAddr, (std::time::Instant, u32)>>,  // IP → (ikkuna-alku, pyyntömäärä)
@@ -260,6 +261,7 @@ async fn main() {
         ip_connections: Mutex::new(HashMap::new()),
         node_ips: Mutex::new(HashMap::new()),
         node_tasks: Mutex::new(HashMap::new()),
+        node_types: Mutex::new(HashMap::new()),
         node_busy: Mutex::new(std::collections::HashSet::new()),
         pending_task_ids: Mutex::new(std::collections::HashSet::new()),
         api_rate_limits: Mutex::new(HashMap::new()),
@@ -382,6 +384,9 @@ async fn main() {
         .route("/api/pairs", get(api_pairs))
         .route("/api/stats", get(api_stats))
         .route("/api/v1/chat/completions", axum::routing::post(api_chat_completions))
+        .route("/api/v1/model", axum::routing::post(api_change_model))
+        .route("/api/v1/hardware", get(api_hardware))
+        .route("/api/v1/ollama/tags", get(api_ollama_tags))
         .route("/admin", get(admin_page))
         .nest_service("/", {
             let static_dir = std::env::var("STATIC_DIR").unwrap_or_else(|_| "../static".to_string());
@@ -677,6 +682,7 @@ async fn handle_socket(socket: WebSocket, state: Arc<AppState>, ip: IpAddr) {
                     state.db.insert_session(node_id, &ip.to_string(), node_type, &json);
                 }
                 state.node_tasks.lock().unwrap().insert(node_id, selected_task);
+                state.node_types.lock().unwrap().insert(node_id, node_type.to_string());
 
                 if node_type == "native" {
                     let sys = json.get("system");
@@ -934,6 +940,7 @@ async fn handle_socket(socket: WebSocket, state: Arc<AppState>, ip: IpAddr) {
         ips.remove(&node_id);
         vram.remove(&node_id);
     }
+    state.node_types.lock().unwrap().remove(&node_id);
     tracing::info!("Solmu {} ({}) poistui verkosta.", node_id, ip);
     broadcast_stats(&state).await;
     sender_task.abort();
@@ -943,6 +950,8 @@ struct ChatCompletionRequest {
     model: String,
     prompt: String,
     task_id: String,
+    #[serde(default)]
+    max_tokens: Option<u64>,
 }
 
 #[derive(serde::Serialize)]
@@ -952,6 +961,78 @@ struct ChatCompletionResponse {
     tokens_generated: u64,
 }
 
+async fn api_ollama_tags() -> axum::response::Response {
+    let ollama_url = std::env::var("OLLAMA_URL").unwrap_or_else(|_| "http://ollama:11434".to_string());
+    match reqwest::get(format!("{}/api/tags", ollama_url)).await {
+        Ok(resp) => {
+            if let Ok(body) = resp.json::<serde_json::Value>().await {
+                axum::Json(body).into_response()
+            } else {
+                axum::Json(serde_json::json!({ "models": [] })).into_response()
+            }
+        }
+        Err(_) => axum::Json(serde_json::json!({ "models": [] })).into_response(),
+    }
+}
+
+async fn api_hardware(
+    axum::extract::State(state): axum::extract::State<Arc<AppState>>,
+) -> axum::response::Response {
+    // Etsitään natiivisolmun GPU-tiedot sessiosta
+    let sessions = state.db.get_sessions(50);
+    let native = sessions.iter().find(|s| {
+        s.get("node_type").and_then(|v| v.as_str()) == Some("native")
+    });
+
+    let (mut vram_mb, mut gpu_name, ram_mb) = if let Some(s) = native {
+        let gpus = s.get("gpus").and_then(|v| v.as_array());
+        let gpu = gpus.and_then(|g| g.first());
+        let vram = gpu.and_then(|g| g.get("vram_total_mb")).and_then(|v| v.as_u64()).unwrap_or(0);
+        let name = gpu.and_then(|g| g.get("name")).and_then(|v| v.as_str()).unwrap_or("").to_string();
+        let ram = s.get("system").and_then(|v| v.get("ram_total_mb")).and_then(|v| v.as_u64()).unwrap_or(0);
+        (vram, name, ram)
+    } else {
+        (0, String::new(), 0)
+    };
+
+    // Fallback: kysytään Ollamalta onko malleja ladattu (= Ollama on käynnissä)
+    if vram_mb == 0 {
+        let ollama_url = std::env::var("OLLAMA_URL").unwrap_or_else(|_| "http://ollama:11434".to_string());
+        if let Ok(resp) = reqwest::get(format!("{}/api/tags", ollama_url)).await {
+            if let Ok(body) = resp.json::<serde_json::Value>().await {
+                let models = body["models"].as_array().map(|a| a.len()).unwrap_or(0);
+                if models > 0 {
+                    gpu_name = "Ollama (GPU/CPU)".to_string();
+                    // Natiivisolmun RAM fallbackina
+                    vram_mb = if ram_mb > 0 { ram_mb } else { 0 };
+                }
+            }
+        }
+    }
+
+    if gpu_name.is_empty() { gpu_name = "ei natiivisolmua".to_string(); }
+
+    axum::Json(serde_json::json!({
+        "gpu_name": gpu_name,
+        "vram_mb": vram_mb,
+        "ram_mb": ram_mb,
+    })).into_response()
+}
+
+async fn api_change_model(
+    axum::extract::State(state): axum::extract::State<Arc<AppState>>,
+    axum::Json(payload): axum::Json<serde_json::Value>,
+) -> axum::response::Response {
+    let model = payload.get("model").and_then(|v| v.as_str()).unwrap_or("");
+    if model.is_empty() {
+        return (axum::http::StatusCode::BAD_REQUEST, "model puuttuu").into_response();
+    }
+    tracing::info!("Mallin vaihto: {}", model);
+    let msg = serde_json::json!({ "type": "change_model", "model": model });
+    let _ = state.stats_tx.send(msg.to_string());
+    axum::Json(serde_json::json!({ "status": "ok", "model": model })).into_response()
+}
+
 async fn api_chat_completions(
     axum::extract::State(state): axum::extract::State<Arc<AppState>>,
     ConnectInfo(addr): ConnectInfo<SocketAddr>,
@@ -966,16 +1047,17 @@ async fn api_chat_completions(
             *entry = (now, 1); // Uusi ikkuna
         } else {
             entry.1 += 1;
-            if entry.1 > 10 {
+            if entry.1 > 30 {
                 return (axum::http::StatusCode::TOO_MANY_REQUESTS, "Liian monta pyyntöä — yritä minuutin kuluttua").into_response();
             }
         }
     }
 
-    // Etsitään vapaa tai varattu solmu, joka vastaa pyydettyä mallia
+    // Etsitään vapaa solmu — priorisoidaan natiivisolmut (GPU) selaimen edelle
     let (target_node_free, target_node_any, total_matching) = {
         let tasks = state.node_tasks.lock().unwrap();
         let busy = state.node_busy.lock().unwrap();
+        let node_types = state.node_types.lock().unwrap();
         let matching: Vec<u64> = tasks.iter().filter(|(_, task)| {
             if payload.model == "qwen-coder" {
                 task.starts_with("qwen-coder")
@@ -983,7 +1065,12 @@ async fn api_chat_completions(
                 **task == payload.model
             }
         }).map(|(k, _)| *k).collect();
-        let free = matching.iter().find(|id| !busy.contains(id)).copied();
+        // Vapaat solmut: natiivi ensin, sitten selain
+        let free_native = matching.iter().find(|id| {
+            !busy.contains(id) && node_types.get(id).map(|t| t == "native").unwrap_or(false)
+        }).copied();
+        let free_any = matching.iter().find(|id| !busy.contains(id)).copied();
+        let free = free_native.or(free_any);
         let any = matching.first().copied();
         (free, any, matching.len())
     };
@@ -1059,12 +1146,15 @@ async fn api_chat_completions(
     state.node_busy.lock().unwrap().insert(target_node_id);
     state.pending_task_ids.lock().unwrap().insert(payload.task_id.clone());
 
-    let msg = serde_json::json!({
+    let mut msg = serde_json::json!({
         "type": "llm_prompt",
         "prompt": payload.prompt,
         "model": payload.model,
         "task_id": payload.task_id,
     });
+    if let Some(mt) = payload.max_tokens {
+        msg.as_object_mut().unwrap().insert("max_tokens".to_string(), serde_json::json!(mt));
+    }
 
     // Odotuskanava valmiiksi (solmu palauttaa tuloksen stats_tx kautta)
     let mut rx = state.stats_tx.subscribe();
@@ -1080,7 +1170,7 @@ async fn api_chat_completions(
         }
     }
 
-    let timeout = tokio::time::timeout(std::time::Duration::from_secs(120), async move {
+    let timeout = tokio::time::timeout(std::time::Duration::from_secs(600), async move {
         loop {
             let msg_str = match rx.recv().await {
                 Ok(msg) => msg,

network-poc/native-node/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "native-node"
-version = "0.1.0"
+version = "0.2.2"
 edition = "2024"
 
 [dependencies]
@@ -12,10 +12,6 @@ serde_json = "1.0"
 sysinfo = "0.30"
 nvml-wrapper = "0.10"
 wgpu = "24"
-candle-core = { version = "0.8", features = ["cuda"] }
+reqwest = { version = "0.12", features = ["json"] }
-candle-nn = "0.8"
-candle-transformers = "0.8"
-hf-hub = "0.4"
-tokenizers = "0.19"
 tracing = "0.1"
 tracing-subscriber = { version = "0.3", features = ["env-filter"] }

network-poc/native-node/src/inference.rs

@@ -1,261 +1,144 @@
-use candle_core::{Device, Tensor, DType};
-use candle_nn::VarBuilder;
-use candle_transformers::models::qwen2::{Config as QwenConfig, ModelForCausalLM as QwenModel};
-use hf_hub::{api::sync::Api, Repo, RepoType};
 use std::time::Instant;
-
+use std::cell::RefCell;
-/// Top-k sampling with temperature and repetition penalty
-fn sample_top_k(logits: &Tensor, k: usize, temperature: f64, generated_tokens: &[u32], repetition_penalty: f64, rng_state: &mut u64) -> Result<u32, String> {
-    let mut logits_vec: Vec<f32> = logits.to_vec1::<f32>().map_err(|e| format!("to_vec1: {}", e))?;
-    if logits_vec.is_empty() { return Err("Tyhjä logits".to_string()); }
-
-    // Repetition penalty: rankaisee jo generoituja tokeneita
-    for &token_id in generated_tokens {
-        if (token_id as usize) < logits_vec.len() {
-            let logit = &mut logits_vec[token_id as usize];
-            if *logit > 0.0 {
-                *logit /= repetition_penalty as f32;
-            } else {
-                *logit *= repetition_penalty as f32;
-            }
-        }
-    }
-
-    // Temperature scaling
-    if temperature > 0.0 && temperature != 1.0 {
-        for logit in logits_vec.iter_mut() {
-            *logit /= temperature as f32;
-        }
-    }
-
-    // Top-k: etsitään k suurinta
-    let mut indexed: Vec<(usize, f32)> = logits_vec.iter().enumerate().map(|(i, &v)| (i, v)).collect();
-    indexed.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal));
-    indexed.truncate(k);
-
-    if k == 1 || temperature == 0.0 {
-        return Ok(indexed[0].0 as u32);
-    }
-
-    // Softmax top-k:lle
-    let max_logit = indexed[0].1;
-    let exps: Vec<f32> = indexed.iter().map(|x| (x.1 - max_logit).exp()).collect();
-    let sum: f32 = exps.iter().sum();
-    let probs: Vec<f32> = exps.iter().map(|e| e / sum).collect();
-
-    // XorShift64 RNG
-    *rng_state ^= *rng_state << 13;
-    *rng_state ^= *rng_state >> 7;
-    *rng_state ^= *rng_state << 17;
-    let rand_val = (*rng_state % 10000) as f32 / 10000.0;
-
-    let mut cumulative = 0.0;
-    for (i, p) in probs.iter().enumerate() {
-        cumulative += p;
-        if rand_val < cumulative {
-            return Ok(indexed[i].0 as u32);
-        }
-    }
-
-    Ok(indexed[0].0 as u32)
-}
 
 pub struct LlmEngine {
-    tokenizer: tokenizers::Tokenizer,
+    ollama_url: String,
-    model: QwenModel,
+    model: RefCell<String>,
-    device: Device,
+    client: reqwest::Client,
-    eos_token: u32,
 }
 
 impl LlmEngine {
-    pub fn load() -> Result<Self, String> {
+    pub async fn load() -> Result<Self, String> {
-        let device = Device::cuda_if_available(0).map_err(|e| format!("Device: {}", e))?;
+        let model = std::env::var("OLLAMA_MODEL").unwrap_or_else(|_| "qwen2.5-coder:7b".to_string());
-        let device_name = if device.is_cuda() { "CUDA" } else { "CPU" };
-        tracing::info!("LLM device: {}", device_name);
 
-        let dtype = if device.is_cuda() { DType::F16 } else { DType::F32 };
+        let client = reqwest::Client::builder()
+            .timeout(std::time::Duration::from_secs(600))
+            .connect_timeout(std::time::Duration::from_secs(3))
+            .build()
+            .map_err(|e| format!("HTTP client: {}", e))?;
 
-        tracing::info!("Ladataan Qwen2.5-Coder-0.5B-Instruct...");
+        // Jos OLLAMA_URL on asetettu, käytetään sitä suoraan
-        let api = Api::new().map_err(|e| format!("HF API: {}", e))?;
+        let ollama_url = if let Ok(url) = std::env::var("OLLAMA_URL") {
-        let repo = api.repo(Repo::with_revision(
+            tracing::info!("Ollama backend (env): {}", url);
-            "Qwen/Qwen2.5-Coder-0.5B-Instruct".to_string(),
+            url
-            RepoType::Model,
-            "main".to_string(),
-        ));
-
-        let tokenizer_path = repo.get("tokenizer.json").map_err(|e| format!("Tokenizer lataus: {}", e))?;
-        let model_path = repo.get("model.safetensors").map_err(|e| format!("Malli lataus: {}", e))?;
-
-        tracing::info!("Ladataan tokenizer: {:?}", tokenizer_path);
-        let tokenizer = tokenizers::Tokenizer::from_file(&tokenizer_path)
-            .map_err(|e| format!("Tokenizer: {}", e))?;
-
-        let config = QwenConfig {
-            vocab_size: 151936,
-            hidden_size: 896,
-            intermediate_size: 4864,
-            num_hidden_layers: 24,
-            num_attention_heads: 14,
-            num_key_value_heads: 2,
-            max_position_embeddings: 32768,
-            sliding_window: 32768,
-            max_window_layers: 21,
-            tie_word_embeddings: true,
-            rope_theta: 1000000.0,
-            rms_norm_eps: 1e-6,
-            use_sliding_window: false,
-            hidden_act: candle_nn::Activation::Silu,
-        };
-
-        let start = Instant::now();
-        let vb = unsafe {
-            VarBuilder::from_mmaped_safetensors(&[model_path.clone()], dtype, &device)
-                .map_err(|e| format!("VarBuilder: {}", e))?
-        };
-        let model = QwenModel::new(&config, vb).map_err(|e| format!("Malli: {}", e))?;
-        tracing::info!("Malli ladattu ({:.1}s) — {}", start.elapsed().as_secs_f64(), device_name);
-
-        Ok(LlmEngine {
-            tokenizer,
-            model,
-            device,
-            eos_token: 151645,
-        })
-    }
-
-    pub fn generate(&mut self, prompt: &str, max_tokens: usize) -> Result<GenerateResult, String> {
-        // Prefill: aloitetaan vastaus ```-koodiblokkilla → malli jatkaa suoraan koodilla
-        let formatted = format!("<|im_start|>system\nYou are a coding assistant. Respond with ONLY code. No explanations, no markdown, no comments unless asked.<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n```\n", prompt);
-
-        let encoding = self.tokenizer.encode(formatted.as_str(), true)
-            .map_err(|e| format!("Encode: {}", e))?;
-        let input_ids: Vec<u32> = encoding.get_ids().to_vec();
-        let input_len = input_ids.len();
-
-        // Nollataan KV-cache edellisestä promptista
-        self.model.clear_kv_cache();
-
-        // Sampling-parametrit
-        let temperature = 0.7;
-        let top_k = 40;
-        let repetition_penalty = 1.15;
-        let mut rng_state: u64 = std::time::SystemTime::now()
-            .duration_since(std::time::UNIX_EPOCH)
-            .unwrap()
-            .as_nanos() as u64;
-
-        let start = Instant::now();
-
-        // Prefill
-        let input = Tensor::new(input_ids.as_slice(), &self.device)
-            .and_then(|t| t.unsqueeze(0))
-            .map_err(|e| format!("Tensor: {}", e))?;
-
-        let logits = self.model.forward(&input, 0)
-            .map_err(|e| format!("Forward prefill: {}", e))?;
-
-        let logits = logits.squeeze(0).map_err(|e| format!("Squeeze: {}", e))?;
-        let logits = if logits.dims().len() == 2 {
-            let seq_len = logits.dim(0).map_err(|e| format!("Dim: {}", e))?;
-            if seq_len == 0 { return Err("Tyhjä tensori".to_string()); }
-            logits.get(seq_len - 1).map_err(|e| format!("Get: {}", e))?
         } else {
-            logits
+            // Haistellaan Ollamaa tunnetuista osoitteista
-        };
+            let candidates = [
-
+                "http://localhost:11434",
-        let mut generated_text = String::new();
+                "http://127.0.0.1:11434",
-        let mut tokens_generated: usize = 0;
+                "http://ollama:11434",
-        let mut all_tokens: Vec<u32> = Vec::new();
+                "http://host.docker.internal:11434",
-
+            ];
-        let mut next_token = sample_top_k(&logits, top_k, temperature, &all_tokens, repetition_penalty, &mut rng_state)?;
+            let mut found = None;
-
+            for url in &candidates {
-        if next_token != self.eos_token {
+                let probe = reqwest::Client::builder()
-            if let Ok(text) = self.tokenizer.decode(&[next_token], true) {
+                    .connect_timeout(std::time::Duration::from_secs(2))
-                generated_text.push_str(&text);
+                    .build().unwrap_or(client.clone());
-            }
+                if let Ok(resp) = probe.get(format!("{}/api/version", url)).send().await {
-            all_tokens.push(next_token);
+                    if resp.status().is_success() {
-            tokens_generated += 1;
+                        tracing::info!("Ollama löytyi osoitteesta: {}", url);
-        }
+                        found = Some(url.to_string());
-
+                        break;
-        // Autoregressive
-        let mut pos = input_len;
-        for _ in 1..max_tokens {
-            if next_token == self.eos_token { break; }
-
-            let input = Tensor::new(&[next_token], &self.device)
-                .and_then(|t| t.unsqueeze(0))
-                .map_err(|e| format!("Tensor: {}", e))?;
-
-            let logits = self.model.forward(&input, pos)
-                .map_err(|e| format!("Forward pos {}: {}", pos, e))?;
-
-            let logits = logits.squeeze(0).map_err(|e| format!("Squeeze: {}", e))?;
-            let logits = if logits.dims().len() == 2 {
-                let seq_len = logits.dim(0).map_err(|e| format!("Dim: {}", e))?;
-                if seq_len == 0 { break; }
-                logits.get(seq_len - 1).map_err(|e| format!("Get: {}", e))?
-            } else {
-                logits
-            };
-            next_token = sample_top_k(&logits, top_k, temperature, &all_tokens, repetition_penalty, &mut rng_state)?;
-            pos += 1;
-
-            if next_token == self.eos_token { break; }
-
-            if let Ok(text) = self.tokenizer.decode(&[next_token], true) {
-                generated_text.push_str(&text);
-
-                // Stop-sekvenssit: katkaistaan kun malli alkaa selittää
-                let lower = generated_text.to_lowercase();
-                if lower.contains("\n###") || lower.contains("\nexplanation") || lower.contains("\nnote:") || lower.contains("\noutput:") || lower.contains("\n```\n\n") || lower.contains("\n// example") || lower.contains("\n# example") {
-                    for stop in &["\n###", "\nExplanation", "\nNote:", "\nOutput:", "\n```\n\n", "\n// Example", "\n// example", "\n# Example", "\n# example"] {
-                        if let Some(pos) = generated_text.find(stop) {
-                            generated_text.truncate(pos);
-                        }
                     }
-                    break;
                 }
             }
-            all_tokens.push(next_token);
+            found.unwrap_or_else(|| {
-            tokens_generated += 1;
+                tracing::warn!("Ollamaa ei löytynyt — käytetään oletusta http://localhost:11434");
+                "http://localhost:11434".to_string()
+            })
+        };
+
+        tracing::info!("Ollama backend: {} | malli: {}", ollama_url, model);
+        Ok(LlmEngine { ollama_url, model: RefCell::new(model), client })
+    }
+
+    pub fn model_name(&self) -> String {
+        self.model.borrow().clone()
+    }
+
+    pub fn set_model(&self, new_model: String) {
+        *self.model.borrow_mut() = new_model;
+    }
+
+    /// Varmistaa että malli on ladattu Ollamaan (ollama pull)
+    pub async fn ensure_model(&self) -> Result<(), String> {
+        let model = self.model.borrow().clone();
+        tracing::info!("Tarkistetaan malli {}...", model);
+        let resp = self.client.post(format!("{}/api/pull", self.ollama_url))
+            .json(&serde_json::json!({ "name": model, "stream": false }))
+            .send()
+            .await
+            .map_err(|e| format!("Ollama pull: {}", e))?;
+
+        if resp.status().is_success() {
+            tracing::info!("Malli {} valmis", model);
+            Ok(())
+        } else {
+            Err(format!("Ollama pull epäonnistui: {}", resp.status()))
+        }
+    }
+
+    pub async fn generate(&self, prompt: &str, max_tokens: usize) -> Result<GenerateResult, String> {
+        let system = "You are a coding assistant. Respond with ONLY code. Use proper newlines and indentation. No explanations, no markdown fences, no comments unless asked.";
+        let model = self.model.borrow().clone();
+
+        let start = Instant::now();
+        let resp = self.client.post(format!("{}/api/generate", self.ollama_url))
+            .json(&serde_json::json!({
+                "model": model,
+                "prompt": prompt,
+                "system": system,
+                "stream": false,
+                "options": {
+                    "num_predict": max_tokens,
+                    "temperature": 0.7,
+                    "top_k": 40,
+                    "repeat_penalty": 1.15,
+                    "stop": ["<|im_end|>", "\n###", "\nExplanation", "\nNote:"]
+                }
+            }))
+            .send()
+            .await
+            .map_err(|e| format!("Ollama generate: {}", e))?;
+
+        if !resp.status().is_success() {
+            return Err(format!("Ollama HTTP {}", resp.status()));
         }
 
-        let gen_time = start.elapsed();
+        let body: serde_json::Value = resp.json().await
-        let tokens_per_sec = if gen_time.as_secs_f64() > 0.0 {
+            .map_err(|e| format!("Ollama JSON: {}", e))?;
-            tokens_generated as f64 / gen_time.as_secs_f64()
+
+        let text = body["response"].as_str().unwrap_or("").to_string();
+        let total_duration_ns = body["total_duration"].as_u64().unwrap_or(0);
+        let eval_count = body["eval_count"].as_u64().unwrap_or(0) as usize;
+        let eval_duration_ns = body["eval_duration"].as_u64().unwrap_or(1);
+
+        let duration_ms = start.elapsed().as_millis() as f64;
+        let tokens_per_sec = if eval_duration_ns > 0 {
+            eval_count as f64 / (eval_duration_ns as f64 / 1_000_000_000.0)
         } else { 0.0 };
 
         Ok(GenerateResult {
-            text: strip_markdown_wrapper(&generated_text),
+            text: strip_code_fences(&text),
-            tokens_generated,
+            tokens_generated: eval_count,
-            duration_ms: gen_time.as_millis() as f64,
+            duration_ms,
             tokens_per_sec,
         })
     }
 }
 
-const LANG_TAGS: &[&str] = &[
+/// Siivoa mahdolliset markdown-koodiblokki-merkit
-    "python", "py", "rust", "rs", "javascript", "js", "typescript", "ts",
+fn strip_code_fences(text: &str) -> String {
-    "java", "kotlin", "scala", "go", "ruby", "rb", "php", "swift",
-    "c", "cpp", "c++", "c#", "csharp", "r", "sql", "bash", "sh", "zsh",
-    "html", "css", "json", "yaml", "yml", "toml", "xml", "markdown", "md",
-    "lua", "perl", "dart", "elixir", "haskell", "hs", "ocaml", "zig",
-    "plaintext", "text", "txt",
-];
-
-/// Siivoa mallin tuottama vastaus (prefill-yhteensopiva).
-fn strip_markdown_wrapper(text: &str) -> String {
     let mut result = text.trim().to_string();
 
-    // 1. Kielitunniste — VAIN tunnettu kieli
+    // Poista aloittava ```lang
-    if let Some(nl) = result.find('\n') {
+    if result.starts_with("```") {
-        let first = result[..nl].trim().to_lowercase();
+        if let Some(nl) = result.find('\n') {
-        if LANG_TAGS.contains(&first.as_str()) {
             result = result[nl + 1..].to_string();
         }
     }
 
-    // 2. Sulkeva ``` — VAIN omalla rivillään lopussa
+    // Poista sulkeva ```
     let trimmed = result.trim_end();
     if trimmed.ends_with("```") {
         let before = &trimmed[..trimmed.len() - 3];
@@ -264,29 +147,7 @@ fn strip_markdown_wrapper(text: &str) -> String {
         }
     }
 
-    // 3. Johdantolauseet
+    result
-    let lower = result.trim().to_lowercase();
-    for prefix in &["sure!", "here is", "here's", "certainly!", "below is"] {
-        if lower.starts_with(prefix) {
-            if let Some(nl) = result.find('\n') { result = result[nl + 1..].to_string(); }
-            break;
-        }
-    }
-
-    // 4. Selityskommentit alusta
-    let mut lines: Vec<&str> = result.trim().lines().collect();
-    while !lines.is_empty() {
-        let first = lines[0].trim();
-        let is_preamble = first.starts_with("# ") && !first.starts_with("#!")
-            && (first.to_lowercase().contains("this is")
-                || first.to_lowercase().contains("simple")
-                || first.to_lowercase().contains("program that")
-                || first.to_lowercase().contains("here is")
-                || first.to_lowercase().contains("the following")
-                || first.to_lowercase().contains("below"));
-        if is_preamble { lines.remove(0); } else { break; }
-    }
-    lines.join("\n").trim().to_string()
 }
 
 pub struct GenerateResult {

network-poc/native-node/src/main.rs

@@ -285,15 +285,19 @@ async fn main() {
         }
     }
 
-    // Ladataan LLM-malli
+    // Ollama-backend
-    tracing::info!("Ladataan LLM-mallia...");
+    tracing::info!("Alustetaan Ollama-yhteyttä...");
-    let mut llm = match inference::LlmEngine::load() {
+    let llm = match inference::LlmEngine::load().await {
         Ok(engine) => {
-            tracing::info!("LLM valmis inferenssiin!");
+            // Varmistetaan malli (ollama pull) — odotetaan kunnes valmis
+            match engine.ensure_model().await {
+                Ok(()) => tracing::info!("Ollama valmis inferenssiin!"),
+                Err(e) => tracing::warn!("Mallin lataus: {} — yritetään silti", e),
+            }
             Some(engine)
         }
         Err(e) => {
-            tracing::warn!("LLM-lataus epäonnistui: {} — toimitaan ilman inferenssiä", e);
+            tracing::warn!("Ollama-alustus epäonnistui: {} — toimitaan ilman inferenssiä", e);
             None
         }
     };
@@ -324,11 +328,13 @@ async fn main() {
                                 
                                 if !prompt.is_empty() && msg_model.starts_with("qwen-coder") {
 
-                                    if let Some(ref mut engine) = llm {
+                                    if let Some(ref engine) = llm {
                                         busy = true;
-                                        tracing::info!("Generoidaan (task_id: {}): \"{}\"", task_id, prompt);
+                                        let max_tokens = task.get("max_tokens").and_then(|v| v.as_u64()).unwrap_or(512) as usize;
+                                        tracing::info!("Generoidaan (task_id: {}, max_tokens: {}): \"{}\"", task_id, max_tokens, &prompt[..prompt.len().min(100)]);
 
-                                        match engine.generate(prompt, 64) {
+                                        let model_name = engine.model_name();
+                                        match engine.generate(prompt, max_tokens).await {
                                             Ok(result) => {
                                                 tracing::info!(
                                                     "Tulos: {} tokenia | {:.0}ms | {:.1} tok/s | \"{}\"",
@@ -341,7 +347,7 @@ async fn main() {
                                                 let done = json!({
                                                     "type": "llm_done",
                                                     "prompt": prompt,
-                                                    "model": "Qwen2.5-Coder-0.5B (native/GPU)",
+                                                    "model": format!("{} (Ollama)", model_name),
                                                     "response": result.text,
                                                     "tokens_generated": result.tokens_generated,
                                                     "duration_ms": result.duration_ms,
@@ -360,7 +366,21 @@ async fn main() {
                                 }
                             }
                         }
-                        // Ohitetaan pair_task, stats jne.
+                        // Mallin vaihto lennossa
+                        if text.contains("change_model") {
+                            if let Ok(task) = serde_json::from_str::<serde_json::Value>(&text) {
+                                if let Some(new_model) = task.get("model").and_then(|v| v.as_str()) {
+                                    if let Some(ref engine) = llm {
+                                        tracing::info!("Vaihdetaan malli: {}", new_model);
+                                        engine.set_model(new_model.to_string());
+                                        match engine.ensure_model().await {
+                                            Ok(()) => tracing::info!("Malli {} valmis!", new_model),
+                                            Err(e) => tracing::error!("Mallin lataus epäonnistui: {}", e),
+                                        }
+                                    }
+                                }
+                            }
+                        }
                     }
                 }
                 tracing::warn!("Yhteys hubiin katkesi — yritetään uudelleen 5s...");

network-poc/node/src/lib.rs

@@ -38,17 +38,50 @@ pub fn set_gpu_load(load: u32) {
     console_log!("[Wasm] GPU Kuormitusraja vaihdettu -> {}%", load);
 }
 
-// Asynkroninen odotus WebAssemblylle
+// Worker-yhteensopiva setTimeout — toimii sekä Window- että Worker-kontekstissa
-async fn sleep_ms(ms: i32) {
+#[wasm_bindgen]
+extern "C" {
+    #[wasm_bindgen(js_name = setTimeout)]
+    fn set_timeout(closure: &js_sys::Function, ms: i32);
+}
+
+// Asynkroninen odotus WebAssemblylle (Window + Worker)
+pub async fn sleep_ms(ms: i32) {
     let promise = js_sys::Promise::new(&mut |resolve, _| {
-        web_sys::window()
+        set_timeout(&resolve, ms);
-            .unwrap()
-            .set_timeout_with_callback_and_timeout_and_arguments_0(&resolve, ms)
-            .unwrap();
     });
     let _ = wasm_bindgen_futures::JsFuture::from(promise).await;
 }
 
+// Worker-yhteensopiva Performance — käyttää globalThis.performance
+pub fn perf_now() -> f64 {
+    js_sys::Reflect::get(&js_sys::global(), &"performance".into())
+        .ok()
+        .and_then(|p| js_sys::Reflect::get(&p, &"now".into()).ok())
+        .and_then(|f| f.dyn_into::<js_sys::Function>().ok())
+        .and_then(|f| {
+            let perf = js_sys::Reflect::get(&js_sys::global(), &"performance".into()).unwrap();
+            f.call0(&perf).ok()
+        })
+        .and_then(|v| v.as_f64())
+        .unwrap_or(0.0)
+}
+
+// Worker-yhteensopiva fetch — käyttää globalThis.fetch
+pub async fn worker_fetch(url: &str) -> Result<web_sys::Response, String> {
+    let promise = js_sys::Reflect::get(&js_sys::global(), &"fetch".into())
+        .map_err(|_| "fetch ei saatavilla".to_string())?
+        .dyn_into::<js_sys::Function>()
+        .map_err(|_| "fetch ei funktio".to_string())?
+        .call1(&JsValue::NULL, &url.into())
+        .map_err(|e| format!("fetch: {:?}", e))?;
+    let resp = wasm_bindgen_futures::JsFuture::from(js_sys::Promise::from(promise))
+        .await
+        .map_err(|e| format!("fetch await: {:?}", e))?;
+    resp.dyn_into::<web_sys::Response>()
+        .map_err(|_| "ei Response".to_string())
+}
+
 // Geneerinen tensorilaskenta — toimii millä tahansa Burn-backendillä
 fn run_matmul<B: burn::tensor::backend::Backend>(size: usize) -> String {
     let device = Default::default();
@@ -85,6 +118,27 @@ async fn run_ai_tensor_inference(difficulty: usize) -> String {
     format!("PoC {} Matmul ({}x{}) >> {}", backend_name, active_workload_size, active_workload_size, result)
 }
 
+/// JS-exportti: tokenisoi tekstin ja palauttaa JSON-merkkijonon
+/// Tokenizer ladataan IndexedDB:stä (täytyy olla ladattu aiemmin)
+#[wasm_bindgen]
+pub async fn tokenize_js(text: String) -> Result<String, JsValue> {
+    let cached_tok = storage::load_from_idb("tokenizer.json").await.unwrap_or(None);
+    let Some(bytes) = cached_tok else {
+        // Yritetään ladata verkosta
+        let resp = reqwest::get("https://huggingface.co/Qwen/Qwen2.5-Coder-0.5B/resolve/main/tokenizer.json").await
+            .map_err(|e| JsValue::from_str(&format!("Tokenizer-lataus epäonnistui: {}", e)))?;
+        let bytes = resp.bytes().await
+            .map_err(|e| JsValue::from_str(&format!("Tokenizer-lataus epäonnistui: {}", e)))?;
+        let _ = storage::save_to_idb("tokenizer.json", &bytes).await;
+        let tokenizer = tokenizers::Tokenizer::from_bytes(&bytes)
+            .map_err(|e| JsValue::from_str(&format!("Tokenizer-parsinta: {}", e)))?;
+        return Ok(tokenize_text(&tokenizer, &text).to_string());
+    };
+    let tokenizer = tokenizers::Tokenizer::from_bytes(&bytes)
+        .map_err(|e| JsValue::from_str(&format!("Tokenizer-parsinta: {}", e)))?;
+    Ok(tokenize_text(&tokenizer, &text).to_string())
+}
+
 /// Tokenisoi yhden tekstin ja palauttaa metriikat
 fn tokenize_text(tokenizer: &tokenizers::Tokenizer, text: &str) -> serde_json::Value {
     let char_count = text.chars().count();
@@ -123,10 +177,9 @@ async fn run_single_tokenize(text: String, ws: Rc<RefCell<WebSocket>>) {
     let Some(bytes) = cached_tok else { return; };
     let Ok(tokenizer) = tokenizers::Tokenizer::from_bytes(&bytes) else { return; };
 
-    let perf = web_sys::window().unwrap().performance().unwrap();
+    let start = perf_now();
-    let start = perf.now();
     let result = tokenize_text(&tokenizer, &text);
-    let duration_ms = perf.now() - start;
+    let duration_ms = perf_now() - start;
 
     let token_count = result["token_count"].as_u64().unwrap_or(0);
     let cpt = result["chars_per_token"].as_f64().unwrap_or(0.0);
@@ -157,11 +210,10 @@ async fn run_pair_comparison(en_text: String, fi_text: String, ws: Rc<RefCell<We
         return;
     };
 
-    let perf = web_sys::window().unwrap().performance().unwrap();
+    let start_time = perf_now();
-    let start_time = perf.now();
     let en_result = tokenize_text(&tokenizer, &en_text);
     let fi_result = tokenize_text(&tokenizer, &fi_text);
-    let duration_ms = perf.now() - start_time; // millisekunteja desimaalitarkkuudella
+    let duration_ms = perf_now() - start_time;
 
     let en_cpt = en_result["chars_per_token"].as_f64().unwrap_or(0.0);
     let fi_cpt = fi_result["chars_per_token"].as_f64().unwrap_or(0.0);

network-poc/node/src/qwen.rs

@@ -24,10 +24,7 @@ async fn ensure_cached(key: &str, url: &str, ws: &Rc<RefCell<WebSocket>>) -> Res
 
     console_log!("[Qwen] Ladataan {}...", key);
 
-    let window = web_sys::window().unwrap();
+    let resp = crate::worker_fetch(url).await?;
-    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".to_string())?;
     if !resp.ok() { return Err(format!("HTTP {}", resp.status())); }
 
     let total_size: usize = resp.headers()
@@ -71,7 +68,7 @@ async fn ensure_cached(key: &str, url: &str, ws: &Rc<RefCell<WebSocket>>) -> Res
 }
 
 pub async fn run_qwen_inference(prompt: String, ws: Rc<RefCell<WebSocket>>) {
-    let perf = web_sys::window().unwrap().performance().unwrap();
+    // performance via crate::perf_now()
 
     let tok_bytes = match ensure_cached("qwen05b-tokenizer.json", TOKENIZER_URL, &ws).await {
         Ok(b) => b,
@@ -88,7 +85,7 @@ pub async fn run_qwen_inference(prompt: String, ws: Rc<RefCell<WebSocket>>) {
     };
 
     console_log!("[Qwen] Rakennetaan mallia...");
-    let start_load = perf.now();
+    let start_load = crate::perf_now();
     let device = Device::Cpu;
     let dtype = DType::F32;
 
@@ -120,7 +117,7 @@ pub async fn run_qwen_inference(prompt: String, ws: Rc<RefCell<WebSocket>>) {
         Err(e) => { console_log!("[Qwen] Mallin lataus: {}", e); return; }
     };
 
-    let load_time = perf.now() - start_load;
+    let load_time = crate::perf_now() - start_load;
     console_log!("[Qwen] Malli ladattu ({:.0}ms). Generoidaan...", load_time);
 
     let encoding = match tokenizer.encode(prompt.as_str(), true) {
@@ -131,7 +128,7 @@ pub async fn run_qwen_inference(prompt: String, ws: Rc<RefCell<WebSocket>>) {
     let input_len = input_ids.len();
     console_log!("[Qwen] 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;
@@ -202,7 +199,7 @@ pub async fn run_qwen_inference(prompt: String, ws: Rc<RefCell<WebSocket>>) {
         crate::sleep_ms(0).await;
     }
 
-    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 };
     console_log!("[Qwen] {} tokenia | {:.0}ms | {:.1} tok/s", tokens_generated, gen_time, tokens_per_sec);
 

network-poc/node/src/qwen_coder.rs

@@ -140,10 +140,7 @@ async fn ensure_cached(key: &str, url: &str, ws: &Rc<RefCell<WebSocket>>) -> Res
 
     console_log!("[Coder] Ladataan {}...", key);
 
-    let window = web_sys::window().unwrap();
+    let resp = crate::worker_fetch(url).await?;
-    let resp_val = wasm_bindgen_futures::JsFuture::from(window.fetch_with_str(url))
-        .await.map_err(|e| format!("Fetch: {:?}", e))?;
-    let resp: web_sys::Response = resp_val.dyn_into().map_err(|_| "Ei Response".to_string())?;
     if !resp.ok() { return Err(format!("HTTP {}", resp.status())); }
 
     let total_size: usize = resp.headers()
@@ -251,17 +248,16 @@ async fn get_or_build_model(use_3b: bool, ws: &Rc<RefCell<WebSocket>>) -> Result
 
 /// use_3b: false = 0.5B (nopea), true = 3B (laadukas)
 pub async fn run_coder_inference(prompt: String, ws: Rc<RefCell<WebSocket>>, use_3b: bool, task_id: Option<String>) {
-    let perf = web_sys::window().unwrap().performance().unwrap();
     let size_label = if use_3b { "3B" } else { "0.5B" };
 
-    let start_load = perf.now();
+    let start_load = crate::perf_now();
 
     if let Err(e) = get_or_build_model(use_3b, &ws).await {
         console_log!("[Coder] Mallin lataus: {}", e);
         return;
     }
 
-    let load_time = perf.now() - start_load;
+    let load_time = crate::perf_now() - start_load;
     if load_time > 100.0 {
         console_log!("[Coder] Malli ladattu ({:.0}ms). Generoidaan...", load_time);
     }
@@ -297,7 +293,7 @@ pub async fn run_coder_inference(prompt: String, ws: Rc<RefCell<WebSocket>>, use
         console_log!("[Coder] Syöte: {} tokenia", input_len);
 
         let device = Device::Cpu;
-        let start_gen = perf.now();
+        let start_gen = crate::perf_now();
         let eos_token = 151645u32;
         let temperature: f32 = 0.7;
         let top_k: usize = 40;
@@ -373,7 +369,7 @@ pub async fn run_coder_inference(prompt: String, ws: Rc<RefCell<WebSocket>>, use
             tokens_generated += 1;
         }
 
-        let gen_time = perf.now() - start_gen;
+        let gen_time = crate::perf_now() - start_gen;
 
         // Siivotaan vastaus: poista markdown-koodiblokit ja johdantotekstit
         let cleaned = strip_markdown_wrapper(&generated_text);

network-poc/node/src/smollm.rs

@@ -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 = crate::worker_fetch(url).await?;
-    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())?;
 
     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!({

network-poc/static/GUIDE.md

@@ -36,29 +36,13 @@ sanan osa, kokonainen sana tai välilyönti. Tokenisaatio tehdään
 BPE-algoritmilla (Byte Pair Encoding) joka oppii yleisimmät
 merkkijonot harjoitusdatasta.
 
-### Esimerkki: koodi
-
-```
-"print('Hello')" → [print] [(' ] [Hello] [')]  = 4 tokenia
-"tulosta('Hei')" → [tul] [osta] [(' ] [He] [i] [')]  = 6 tokenia
-```
-
-Koodi tokenisoidaan tehokkaasti koska `print`, `def`, `return` yms.
-ovat kokonaisia tokeneita. Suomenkielinen `tulosta` joudutaan pilkkomaan
-osiin koska se ei esiinny harjoitusdatassa kokonaisena.
-
 ### Esimerkki: suomi vs. englanti
 
-Sama lause kahdella kielellä Qwen2.5-Coder -tokenisaattorilla:
+Alla oikea tokenisointitulos Qwen2.5-Coder-tokenisaattorilla. Jokainen
+värikoodattu lohko on yksi tokeni — huomaa miten suomi vaatii enemmän
+tokeneita saman merkityksen välittämiseen:
 
-| | Teksti | Tokenit | Määrä | Merkkejä/token |
+![Tokenisointivertailu EN/FI](/images/tokenization-example.png)
-|---|---|---|---|---|
-| EN | The cat sat on the mat | [The] [ cat] [ sat] [ on] [ the] [ mat] | **6** | 3.7 |
-| FI | Kissa istui matolla | [K] [issa] [ ist] [ui] [ mat] [olla] | **6** | 3.2 |
-| EN | Distributed computing in the browser | [Dist] [ributed] [ computing] [ in] [ the] [ browser] | **6** | 6.0 |
-| FI | Hajautettu laskenta selaimessa | [H] [aj] [au] [tettu] [ las] [kenta] [ sel] [aim] [essa] | **9** | 3.3 |
-| EN | Write a function that sorts a list | [Write] [ a] [ function] [ that] [ sorts] [ a] [ list] | **7** | 5.0 |
-| FI | Kirjoita funktio joka lajittelee listan | [K] [irj] [oita] [ funkt] [io] [ joka] [ laj] [ittel] [ee] [ listan] | **10** | 4.0 |
 
 **Huomaa miten:**
 - Englannin yleiset sanat (`the`, `in`, `a`, `function`) ovat kokonaisia tokeneita

network-poc/static/worker.js

@@ -0,0 +1,38 @@
+// Kipinä WASM Worker (ES module) — ajaa kielimallin inferenssin erillisessä säikeessä
+import init, { start_agent_node, set_gpu_load, set_auto_tasks } from './pkg/node.js';
+
+let wasmReady = false;
+
+// Välitetään console.log -viestit pääsäikeelle jotta UI-kuuntelijat näkevät ne
+const _origLog = console.log;
+console.log = function(...args) {
+    _origLog.apply(console, args);
+    self.postMessage({ type: 'log', message: args.join(' ') });
+};
+
+self.onmessage = async (e) => {
+    const { type, data } = e.data;
+
+    if (type === 'init') {
+        try {
+            await init();
+            wasmReady = true;
+            self.postMessage({ type: 'ready' });
+        } catch (err) {
+            self.postMessage({ type: 'error', message: 'WASM init: ' + err.message });
+        }
+    } else if (type === 'start') {
+        if (!wasmReady) return;
+        const { hubUrl, hasWebGPU, deviceInfo, taskId } = data;
+        try {
+            await start_agent_node(hubUrl, hasWebGPU, deviceInfo, taskId);
+            self.postMessage({ type: 'started' });
+        } catch (err) {
+            self.postMessage({ type: 'error', message: 'Node: ' + err.message });
+        }
+    } else if (type === 'set_gpu_load') {
+        if (wasmReady) set_gpu_load(data.load);
+    } else if (type === 'set_auto_tasks') {
+        if (wasmReady) set_auto_tasks(data.enabled);
+    }
+};