on tämä työmaa

network-poc/USER-README.md

@@ -55,8 +55,28 @@ cargo run -p hub
 
 # 3. Avaa selain: http://localhost:3000
 
-# 4. Valinnainen: natiivi-node (terminaali 2)
-HUB_URL=ws://localhost:3000/ws ALLOCATED_GB=4 cargo run -p native-node
+# 4. Valinnainen: natiivi-node LLM-inferenssillä (terminaali 2)
+# Lataa Qwen2.5-0.5B automaattisesti HuggingFacesta (~990 MB, cachetetaan)
+# Release-moodissa ~11 tok/s CPU:lla (32 ydintä)
+CARGO_TARGET_DIR=target-native HUB_URL=ws://localhost:3000/ws ALLOCATED_GB=4 cargo run --release -p native-node
+
+# Tai yhdistä tuotantopalvelimeen:
+CARGO_TARGET_DIR=target-native HUB_URL=wss://kipina.studio/ws ALLOCATED_GB=4 cargo run --release -p native-node
+```
+
+### CUDA-tuki (valinnainen)
+
+Jos koneessa on NVIDIA GPU ja CUDA toolkit:
+
+```bash
+# Asenna CUDA toolkit (Ubuntu/Pop!_OS)
+sudo apt install nvidia-cuda-toolkit
+
+# Muokkaa native-node/Cargo.toml:
+# candle-core = { version = "0.8", features = ["cuda"] }
+
+# Aja — malli käyttää automaattisesti GPU:ta
+CARGO_TARGET_DIR=target-native HUB_URL=ws://localhost:3000/ws cargo run --release -p native-node
 ```
 
 ## WebGPU-asetukset selaimessa

network-poc/hub/src/main.rs

@@ -279,14 +279,32 @@ async fn main() {
                 "What makes Rust special?",
             ];
             let llm_idx = (rng_state as usize / 7) % llm_prompts.len();
-            let llm_msg = serde_json::json!({
+
+            // SmolLM-prompt
+            let smollm_msg = serde_json::json!({
                 "type": "llm_prompt",
                 "prompt": llm_prompts[llm_idx],
                 "model": "smollm-135m",
             });
-            let _ = state_for_task.stats_tx.send(llm_msg.to_string());
+            let _ = state_for_task.stats_tx.send(smollm_msg.to_string());
 
-            tracing::debug!("Tehtävät lähetetty: pair + llm_prompt");
+            // Qwen-prompt (sama prompti, eri malli-tagi)
+            let qwen_msg = serde_json::json!({
+                "type": "llm_prompt",
+                "prompt": llm_prompts[llm_idx],
+                "model": "qwen-05b",
+            });
+            let _ = state_for_task.stats_tx.send(qwen_msg.to_string());
+
+            // Phi-3 prompt
+            let phi3_msg = serde_json::json!({
+                "type": "llm_prompt",
+                "prompt": llm_prompts[llm_idx],
+                "model": "phi3-mini",
+            });
+            let _ = state_for_task.stats_tx.send(phi3_msg.to_string());
+
+            tracing::debug!("Tehtävät lähetetty: pair + smollm + qwen + phi3");
         }
     });
 

network-poc/native-node/Cargo.toml

@@ -12,5 +12,10 @@ serde_json = "1.0"
 sysinfo = "0.30"
 nvml-wrapper = "0.10"
 wgpu = "24"
+candle-core = { version = "0.8" }
+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

@@ -0,0 +1,181 @@
+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::path::PathBuf;
+use std::time::Instant;
+
+pub struct LlmEngine {
+    tokenizer: tokenizers::Tokenizer,
+    model_path: PathBuf,
+    device: Device,
+    dtype: DType,
+    config: QwenConfig,
+    eos_token: u32,
+}
+
+impl LlmEngine {
+    pub fn load() -> Result<Self, String> {
+        let device = Device::cuda_if_available(0).map_err(|e| format!("Device: {}", e))?;
+        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 };
+
+        tracing::info!("Ladataan Qwen2.5-0.5B-Instruct...");
+        let api = Api::new().map_err(|e| format!("HF API: {}", e))?;
+        let repo = api.repo(Repo::with_revision(
+            "Qwen/Qwen2.5-0.5B-Instruct".to_string(),
+            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,
+        };
+
+        // Testi-lataus varmistaa, että painot toimivat
+        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_path,
+            device,
+            dtype,
+            config,
+            eos_token: 151645,
+        })
+    }
+
+    /// Luo tuore malliinstanssi (nollaa KV-cachen)
+    fn fresh_model(&self) -> Result<QwenModel, String> {
+        let vb = unsafe {
+            VarBuilder::from_mmaped_safetensors(&[self.model_path.clone()], self.dtype, &self.device)
+                .map_err(|e| format!("VarBuilder: {}", e))?
+        };
+        QwenModel::new(&self.config, vb).map_err(|e| format!("Malli: {}", e))
+    }
+
+    pub fn generate(&mut self, prompt: &str, max_tokens: usize) -> Result<GenerateResult, String> {
+        let formatted = format!("<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\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();
+
+        // Tuore malli joka promptille (nollaa KV-cachen)
+        let mut model = self.fresh_model()?;
+
+        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 = 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 {
+            logits.get(logits.dim(0).unwrap() - 1).map_err(|e| format!("Get: {}", e))?
+        } else {
+            logits
+        };
+        let mut next_token = logits.argmax(0)
+            .map_err(|e| format!("Argmax: {}", e))?
+            .to_vec0::<u32>()
+            .map_err(|e| format!("to_vec0: {}", e))?;
+
+        let mut generated_text = String::new();
+        let mut tokens_generated: usize = 0;
+        let mut all_tokens: Vec<u32> = Vec::new();
+
+        if next_token != self.eos_token {
+            if let Ok(text) = self.tokenizer.decode(&[next_token], true) {
+                generated_text.push_str(&text);
+            }
+            all_tokens.push(next_token);
+            tokens_generated += 1;
+        }
+
+        // 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 = 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 {
+                logits.get(logits.dim(0).unwrap() - 1).map_err(|e| format!("Get: {}", e))?
+            } else {
+                logits
+            };
+            next_token = logits.argmax(0)
+                .map_err(|e| format!("Argmax: {}", e))?
+                .to_vec0::<u32>()
+                .map_err(|e| format!("to_vec0: {}", e))?;
+            pos += 1;
+
+            if next_token == self.eos_token { break; }
+
+            if let Ok(text) = self.tokenizer.decode(&[next_token], true) {
+                generated_text.push_str(&text);
+            }
+            all_tokens.push(next_token);
+            tokens_generated += 1;
+        }
+
+        let gen_time = start.elapsed();
+        let tokens_per_sec = if gen_time.as_secs_f64() > 0.0 {
+            tokens_generated as f64 / gen_time.as_secs_f64()
+        } else { 0.0 };
+
+        Ok(GenerateResult {
+            text: generated_text,
+            tokens_generated,
+            duration_ms: gen_time.as_millis() as f64,
+            tokens_per_sec,
+        })
+    }
+}
+
+pub struct GenerateResult {
+    pub text: String,
+    pub tokens_generated: usize,
+    pub duration_ms: f64,
+    pub tokens_per_sec: f64,
+}

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

@@ -4,6 +4,8 @@ use sysinfo::System;
 use tokio_tungstenite::connect_async;
 use tokio_tungstenite::tungstenite::Message;
 
+mod inference;
+
 /// GPU-tietorakenne — yhtenäinen kaikille valmistajille
 struct GpuInfo {
     name: String,
@@ -282,7 +284,20 @@ async fn main() {
         }
     }
 
-    // Yhdistetään hubiin — yritetään uudelleen katkon sattuessa
+    // Ladataan LLM-malli
+    tracing::info!("Ladataan LLM-mallia...");
+    let mut llm = match inference::LlmEngine::load() {
+        Ok(engine) => {
+            tracing::info!("LLM valmis inferenssiin!");
+            Some(engine)
+        }
+        Err(e) => {
+            tracing::warn!("LLM-lataus epäonnistui: {} — toimitaan ilman inferenssiä", e);
+            None
+        }
+    };
+
+    // Yhdistetään hubiin
     loop {
         match connect_async(&hub_url).await {
             Ok((ws_stream, _)) => {
@@ -295,17 +310,51 @@ async fn main() {
                     continue;
                 }
 
+                let mut busy = false;
+
                 while let Some(Ok(msg)) = read.next().await {
                     if let Message::Text(text) = msg {
-                        if text.contains("pair_task") || text.contains("ai_task") {
-                            tracing::debug!("Tehtävä vastaanotettu: {}", &text[..text.len().min(80)]);
-                            let reply = json!({
-                                "type": "result",
-                                "status": "success",
-                                "data": "native-node: ei vielä laskentaa"
+                        // LLM-promptit
+                        if text.contains("llm_prompt") && !busy {
+                            if let Ok(task) = serde_json::from_str::<serde_json::Value>(&text) {
+                                let prompt = task.get("prompt").and_then(|v| v.as_str()).unwrap_or("");
+                                if !prompt.is_empty() {
+                                    if let Some(ref mut engine) = llm {
+                                        busy = true;
+                                        tracing::info!("Generoidaan: \"{}\"", prompt);
+
+                                        match engine.generate(prompt, 64) {
+                                            Ok(result) => {
+                                                tracing::info!(
+                                                    "Tulos: {} tokenia | {:.0}ms | {:.1} tok/s | \"{}\"",
+                                                    result.tokens_generated,
+                                                    result.duration_ms,
+                                                    result.tokens_per_sec,
+                                                    &result.text[..result.text.len().min(80)]
+                                                );
+
+                                                let done = json!({
+                                                    "type": "llm_done",
+                                                    "prompt": prompt,
+                                                    "model": "Qwen2.5-0.5B-Instruct (native/GPU)",
+                                                    "response": result.text,
+                                                    "tokens_generated": result.tokens_generated,
+                                                    "duration_ms": result.duration_ms,
+                                                    "tokens_per_sec": (result.tokens_per_sec * 10.0).round() / 10.0,
+                                                    "load_time_ms": 0,
                                                 });
-                            let _ = write.send(Message::Text(reply.to_string())).await;
+                                                let _ = write.send(Message::Text(done.to_string())).await;
                                             }
+                                            Err(e) => {
+                                                tracing::error!("Inferenssivirhe: {}", e);
+                                            }
+                                        }
+                                        busy = false;
+                                    }
+                                }
+                            }
+                        }
+                        // Ohitetaan pair_task, stats jne.
                     }
                 }
                 tracing::warn!("Yhteys hubiin katkesi — yritetään uudelleen 5s...");

network-poc/node/src/lib.rs

@@ -8,6 +8,8 @@ use burn::backend::{Wgpu, NdArray};
 
 pub mod storage;
 pub mod smollm;
+pub mod qwen;
+pub mod phi3;
 
 #[macro_export]
 macro_rules! console_log {
@@ -16,8 +18,9 @@ macro_rules! console_log {
 
 static GPU_LOAD_PERCENT: AtomicU32 = AtomicU32::new(50);
 static HAS_WEBGPU: AtomicBool = AtomicBool::new(true);
-// Valittu tehtävä: 0=tokenize, 1=smollm-135m, 2=qwen-05b, 3=phi3-mini
 static SELECTED_TASK: AtomicU32 = AtomicU32::new(0);
+// Estää rinnakkaiset LLM-inferenssit (vain yksi kerrallaan)
+static LLM_BUSY: AtomicBool = AtomicBool::new(false);
 
 #[wasm_bindgen]
 pub fn set_gpu_load(load: u32) {
@@ -202,14 +205,46 @@ pub async fn start_agent_node(hub_url: String, has_webgpu: bool, device_info_jso
                     }
                 }
             } else if msg.contains("llm_prompt") && current_task == 1 {
-                // Vain SmolLM-solmut käsittelevät llm_prompt-viestejä
-                if let Ok(task) = serde_json::from_str::<serde_json::Value>(&msg) {
+                // Vain SmolLM-solmut, ja vain yksi inferenssi kerrallaan
+                if LLM_BUSY.load(Ordering::SeqCst) {
+                    // Ohitetaan — edellinen inferenssi vielä käynnissä
+                } else if let Ok(task) = serde_json::from_str::<serde_json::Value>(&msg) {
                     let prompt = task.get("prompt").and_then(|v| v.as_str()).unwrap_or("").to_string();
                     let model = task.get("model").and_then(|v| v.as_str()).unwrap_or("").to_string();
                     if !prompt.is_empty() && model == "smollm-135m" {
+                        LLM_BUSY.store(true, Ordering::SeqCst);
                         let ws_for_async = ws_clone.clone();
                         wasm_bindgen_futures::spawn_local(async move {
                             smollm::run_smollm_inference(prompt, ws_for_async).await;
+                            LLM_BUSY.store(false, Ordering::SeqCst);
+                        });
+                    }
+                }
+            } else if msg.contains("llm_prompt") && current_task == 2 {
+                // Qwen2.5-0.5B
+                if LLM_BUSY.load(Ordering::SeqCst) {
+                } else if let Ok(task) = serde_json::from_str::<serde_json::Value>(&msg) {
+                    let prompt = task.get("prompt").and_then(|v| v.as_str()).unwrap_or("").to_string();
+                    if !prompt.is_empty() {
+                        LLM_BUSY.store(true, Ordering::SeqCst);
+                        let ws_for_async = ws_clone.clone();
+                        wasm_bindgen_futures::spawn_local(async move {
+                            qwen::run_qwen_inference(prompt, ws_for_async).await;
+                            LLM_BUSY.store(false, Ordering::SeqCst);
+                        });
+                    }
+                }
+            } else if msg.contains("llm_prompt") && current_task == 3 {
+                // Phi-3 Mini
+                if LLM_BUSY.load(Ordering::SeqCst) {
+                } else if let Ok(task) = serde_json::from_str::<serde_json::Value>(&msg) {
+                    let prompt = task.get("prompt").and_then(|v| v.as_str()).unwrap_or("").to_string();
+                    if !prompt.is_empty() {
+                        LLM_BUSY.store(true, Ordering::SeqCst);
+                        let ws_for_async = ws_clone.clone();
+                        wasm_bindgen_futures::spawn_local(async move {
+                            phi3::run_phi3_inference(prompt, ws_for_async).await;
+                            LLM_BUSY.store(false, Ordering::SeqCst);
                         });
                     }
                 }

network-poc/node/src/phi3.rs

@@ -0,0 +1,36 @@
+use candle_core::{Device, Tensor, DType};
+use candle_nn::VarBuilder;
+use candle_transformers::models::phi3::{Config as Phi3Config, Model as Phi3Model};
+use wasm_bindgen::JsCast;
+use std::cell::RefCell;
+use std::rc::Rc;
+use web_sys::WebSocket;
+
+use crate::storage;
+
+macro_rules! console_log {
+    ($($t:tt)*) => (web_sys::console::log_1(&format_args!($($t)*).to_string().into()))
+}
+
+const MODEL_URL: &str = "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/resolve/main/model.safetensors.index.json";
+const TOKENIZER_URL: &str = "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/resolve/main/tokenizer.json";
+
+// Phi-3 Mini on iso (7.6 GB) — käytetään kvantisoidumpaa versiota myöhemmin
+// Tällä hetkellä: placeholder joka raportoi koon ja jättää inferenssin väliin
+pub async fn run_phi3_inference(prompt: String, ws: Rc<RefCell<WebSocket>>) {
+    console_log!("[Phi-3] Phi-3 Mini 3.8B on liian suuri selaimessa ajettavaksi (~7.6 GB).");
+    console_log!("[Phi-3] Käytä SmolLM 135M tai Qwen2.5 0.5B selaininferenssiin.");
+    console_log!("[Phi-3] Phi-3 tuetaan native-node:lla (Docker + GPU).");
+
+    let done = serde_json::json!({
+        "type": "llm_done",
+        "prompt": prompt,
+        "model": "Phi-3-Mini (ei tuettu selaimessa)",
+        "response": "Phi-3 Mini 3.8B on liian suuri selaimessa ajettavaksi. Käytä SmolLM 135M tai Qwen2.5 0.5B.",
+        "tokens_generated": 0,
+        "duration_ms": 0,
+        "tokens_per_sec": 0,
+        "load_time_ms": 0,
+    });
+    let _ = ws.borrow().send_with_str(&done.to_string());
+}

network-poc/node/src/qwen.rs

@@ -0,0 +1,219 @@
+use candle_core::{Device, Tensor, DType};
+use candle_nn::VarBuilder;
+use candle_transformers::models::qwen2::{Config as QwenConfig, ModelForCausalLM as QwenModel};
+use wasm_bindgen::JsCast;
+use std::cell::RefCell;
+use std::rc::Rc;
+use web_sys::WebSocket;
+
+use crate::storage;
+
+macro_rules! console_log {
+    ($($t:tt)*) => (web_sys::console::log_1(&format_args!($($t)*).to_string().into()))
+}
+
+const MODEL_URL: &str = "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/main/model.safetensors";
+const TOKENIZER_URL: &str = "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer.json";
+
+/// Streaming-lataus HuggingFacesta IndexedDB-cacheen
+async fn ensure_cached(key: &str, url: &str, ws: &Rc<RefCell<WebSocket>>) -> Result<Vec<u8>, String> {
+    if let Ok(Some(bytes)) = storage::load_from_idb(key).await {
+        console_log!("[Qwen] {} löytyi välimuistista ({} MB)", key, bytes.len() / 1024 / 1024);
+        return Ok(bytes);
+    }
+
+    console_log!("[Qwen] Ladataan {}...", key);
+
+    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".to_string())?;
+    if !resp.ok() { return Err(format!("HTTP {}", resp.status())); }
+
+    let total_size: usize = resp.headers()
+        .get("content-length").ok().flatten()
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(0);
+
+    let body = resp.body().ok_or("Ei bodyä")?;
+    let reader: web_sys::ReadableStreamDefaultReader = body.get_reader().dyn_into().map_err(|_| "Ei reader".to_string())?;
+
+    let mut data: Vec<u8> = Vec::with_capacity(total_size);
+    let mut last_pct: u32 = 0;
+
+    loop {
+        let chunk = wasm_bindgen_futures::JsFuture::from(reader.read())
+            .await.map_err(|e| format!("Read: {:?}", e))?;
+        let done = js_sys::Reflect::get(&chunk, &"done".into()).ok().and_then(|v| v.as_bool()).unwrap_or(true);
+        if done { break; }
+        let value = js_sys::Reflect::get(&chunk, &"value".into()).map_err(|_| "value puuttuu".to_string())?;
+        let array = js_sys::Uint8Array::new(&value);
+        let mut buf = vec![0u8; array.length() as usize];
+        array.copy_to(&mut buf);
+        data.extend_from_slice(&buf);
+
+        if total_size > 0 {
+            let pct = ((data.len() as f64 / total_size as f64) * 100.0) as u32;
+            if pct >= last_pct + 5 || pct == 100 {
+                last_pct = pct;
+                console_log!("[Qwen] {} lataus: {}%", key, pct);
+                let msg = serde_json::json!({ "type": "download_progress", "file": key, "pct": pct, "loaded_mb": data.len()/1024/1024, "total_mb": total_size/1024/1024 });
+                let _ = ws.borrow().send_with_str(&msg.to_string());
+            }
+        }
+    }
+
+    console_log!("[Qwen] Tallennetaan {} ({} MB)...", key, data.len() / 1024 / 1024);
+    let _ = storage::save_to_idb(key, &data).await;
+    console_log!("[Qwen] {} tallennettu!", key);
+
+    Ok(data)
+}
+
+pub async fn run_qwen_inference(prompt: String, ws: Rc<RefCell<WebSocket>>) {
+    let perf = web_sys::window().unwrap().performance().unwrap();
+
+    let tok_bytes = match ensure_cached("qwen05b-tokenizer.json", TOKENIZER_URL, &ws).await {
+        Ok(b) => b,
+        Err(e) => { console_log!("[Qwen] Tokenizer-virhe: {}", e); return; }
+    };
+    let tokenizer = match tokenizers::Tokenizer::from_bytes(&tok_bytes) {
+        Ok(t) => t,
+        Err(e) => { console_log!("[Qwen] Tokenizer-parsinta: {}", e); return; }
+    };
+
+    let model_bytes = match ensure_cached("qwen05b-model.safetensors", MODEL_URL, &ws).await {
+        Ok(b) => b,
+        Err(e) => { console_log!("[Qwen] Malli-virhe: {}", e); return; }
+    };
+
+    console_log!("[Qwen] Rakennetaan mallia...");
+    let start_load = perf.now();
+    let device = Device::Cpu;
+    let dtype = DType::F32;
+
+    let tensors = match candle_core::safetensors::load_buffer(&model_bytes, &device) {
+        Ok(t) => t,
+        Err(e) => { console_log!("[Qwen] Safetensors: {}", e); return; }
+    };
+    let vb = VarBuilder::from_tensors(tensors, dtype, &device);
+
+    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 mut model = match QwenModel::new(&config, vb) {
+        Ok(m) => m,
+        Err(e) => { console_log!("[Qwen] Mallin lataus: {}", e); return; }
+    };
+
+    let load_time = perf.now() - start_load;
+    console_log!("[Qwen] Malli ladattu ({:.0}ms). Generoidaan...", load_time);
+
+    let encoding = match tokenizer.encode(prompt.as_str(), true) {
+        Ok(e) => e,
+        Err(e) => { console_log!("[Qwen] Tokenisointivirhe: {}", e); return; }
+    };
+    let input_ids: Vec<u32> = encoding.get_ids().to_vec();
+    let input_len = input_ids.len();
+    console_log!("[Qwen] Syöte: {} tokenia", input_len);
+
+    let start_gen = perf.now();
+    let max_new_tokens = 32;
+    let mut generated_text = String::new();
+    let mut tokens_generated: usize = 0;
+
+    // Prefill
+    let input = match Tensor::new(input_ids.as_slice(), &device).and_then(|t| t.unsqueeze(0)) {
+        Ok(t) => t,
+        Err(e) => { console_log!("[Qwen] Tensor: {}", e); return; }
+    };
+    let logits = match model.forward(&input, 0) {
+        Ok(l) => l,
+        Err(e) => { console_log!("[Qwen] Forward (prefill): {}", e); return; }
+    };
+
+    // Forward palauttaa [batch, vocab_size] tai [batch, seq_len, vocab_size]
+    let logits = logits.squeeze(0).unwrap();
+    let logits = if logits.dims().len() == 2 {
+        // [seq_len, vocab_size] — ota viimeinen
+        logits.get(logits.dim(0).unwrap() - 1).unwrap()
+    } else {
+        logits // jo [vocab_size]
+    };
+    let mut next_token = logits.argmax(0).unwrap().to_vec0::<u32>().unwrap();
+    console_log!("[Qwen] Ensimmäinen token: {}", next_token);
+
+    let eos_token = 151645u32; // <|endoftext|> for Qwen2.5
+
+    if next_token != eos_token {
+        if let Ok(text) = tokenizer.decode(&[next_token], true) {
+            generated_text.push_str(&text);
+            let chunk = serde_json::json!({ "type": "llm_chunk", "token": text, "prompt": prompt, "model": "Qwen2.5-0.5B" });
+            let _ = ws.borrow().send_with_str(&chunk.to_string());
+        }
+        tokens_generated += 1;
+    }
+
+    // Autoregressive
+    let mut pos = input_len;
+    for _ in 1..max_new_tokens {
+        if next_token == eos_token { break; }
+
+        let input = match Tensor::new(&[next_token], &device).and_then(|t| t.unsqueeze(0)) {
+            Ok(t) => t,
+            Err(e) => { console_log!("[Qwen] Tensor: {}", e); break; }
+        };
+        let logits = match model.forward(&input, pos) {
+            Ok(l) => l,
+            Err(e) => { console_log!("[Qwen] Forward pos {}: {}", pos, e); break; }
+        };
+
+        let logits = logits.squeeze(0).unwrap();
+        let logits = if logits.dims().len() == 2 {
+            logits.get(logits.dim(0).unwrap() - 1).unwrap()
+        } else {
+            logits
+        };
+        next_token = logits.argmax(0).unwrap().to_vec0::<u32>().unwrap();
+        pos += 1;
+
+        if next_token == eos_token { break; }
+
+        if let Ok(text) = tokenizer.decode(&[next_token], true) {
+            generated_text.push_str(&text);
+            let chunk = serde_json::json!({ "type": "llm_chunk", "token": text, "prompt": prompt, "model": "Qwen2.5-0.5B" });
+            let _ = ws.borrow().send_with_str(&chunk.to_string());
+        }
+        tokens_generated += 1;
+    }
+
+    let gen_time = 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);
+
+    let done = serde_json::json!({
+        "type": "llm_done",
+        "prompt": prompt,
+        "model": "Qwen2.5-0.5B-Instruct",
+        "response": generated_text,
+        "tokens_generated": tokens_generated,
+        "duration_ms": (gen_time * 100.0).round() / 100.0,
+        "tokens_per_sec": (tokens_per_sec * 10.0).round() / 10.0,
+        "load_time_ms": (load_time * 100.0).round() / 100.0,
+    });
+    let _ = ws.borrow().send_with_str(&done.to_string());
+}

network-poc/node/src/smollm.rs

@@ -1,7 +1,7 @@
 use candle_core::{Device, Tensor, DType};
 use candle_nn::VarBuilder;
 use candle_transformers::models::llama::{Llama, LlamaConfig, LlamaEosToks, Cache};
-use candle_transformers::generation::LogitsProcessor;
+// LogitsProcessor poistettu — käytetään greedy samplingia (argmax) Wasm-yhteensopivuuden vuoksi
 use wasm_bindgen::JsCast;
 use std::cell::RefCell;
 use std::rc::Rc;
@@ -160,8 +160,9 @@ pub async fn run_smollm_inference(prompt: String, ws: Rc<RefCell<WebSocket>>) {
     let load_time = perf.now() - start_load;
     console_log!("[SmolLM] Malli ladattu ({:.0}ms). Generoidaan...", load_time);
 
-    // 3. Tokenisoi syöte
-    let encoding = match tokenizer.encode(prompt.as_str(), true) {
+    // 3. Tokenisoi syöte (Käytetään ChatML-formaattia SmolLM-Instructille)
+    let formatted_prompt = format!("<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n", prompt);
+    let encoding = match tokenizer.encode(formatted_prompt.as_str(), true) {
         Ok(e) => e,
         Err(e) => { console_log!("[SmolLM] Tokenisointivirhe: {}", e); return; }
     };
@@ -172,62 +173,76 @@ pub async fn run_smollm_inference(prompt: String, ws: Rc<RefCell<WebSocket>>) {
 
     // 4. Generoi tokeneita
     let start_gen = perf.now();
-    let mut logits_processor = LogitsProcessor::new(42, Some(0.8), Some(0.95));
-    let mut all_tokens = input_ids.clone();
-    let max_new_tokens = 64;
+    let max_new_tokens = 32;
     let mut generated_text = String::new();
+    let mut tokens_generated: usize = 0;
+    let mut pos: usize = 0;
 
-    for i in 0..max_new_tokens {
-        let context_tokens = if i == 0 {
-            all_tokens.as_slice()
-        } else {
-            std::slice::from_ref(all_tokens.last().unwrap())
+    // Ensimmäinen forward: koko syöte kerralla
+    let input = match Tensor::new(input_ids.as_slice(), &device).and_then(|t| t.unsqueeze(0)) {
+        Ok(t) => t,
+        Err(e) => { console_log!("[SmolLM] Tensor-virhe: {}", e); return; }
     };
 
-        let input = Tensor::new(context_tokens, &device).unwrap().unsqueeze(0).unwrap();
-        let seq_len = input.dim(1).unwrap();
-
-        let logits = match model.forward(&input, input_len + i - seq_len, &mut cache) {
+    let logits = match model.forward(&input, 0, &mut cache) {
         Ok(l) => l,
-            Err(e) => { console_log!("[SmolLM] Forward-virhe stepissä {}: {}", i, e); break; }
+        Err(e) => { console_log!("[SmolLM] Forward-virhe (prefill): {}", e); return; }
     };
 
-        // Viimeisen tokenin logitit
+    // Llama forward voi palauttaa [batch, vocab] tai [batch, seq_len, vocab]
     let logits = logits.squeeze(0).unwrap();
-        let last_dim = logits.dim(0).unwrap();
-        let logits = if last_dim > 1 {
-            logits.get(last_dim - 1).unwrap()
+    let logits = if logits.dims().len() == 2 {
+        logits.get(logits.dim(0).unwrap() - 1).unwrap()
     } else {
-            logits.get(0).unwrap()
+        logits
     };
+    let mut next_token = logits.argmax(0).unwrap().to_vec0::<u32>().unwrap();
+    console_log!("[SmolLM] Ensimmäinen generoitu token: {}", next_token);
+    pos = input_len;
 
-        let next_token = logits_processor.sample(&logits).unwrap();
-
-        // EOS-tarkistus
-        if next_token == 2 {
-            break;
-        }
-
-        all_tokens.push(next_token);
-
-        // Dekoodaa token tekstiksi
+    if next_token != 2 {
         if let Ok(text) = tokenizer.decode(&[next_token], true) {
             generated_text.push_str(&text);
-
-            // Streamaa token hubille
-            let chunk = serde_json::json!({
-                "type": "llm_chunk",
-                "token": text,
-                "is_last": false,
-                "prompt": prompt,
-                "model": "SmolLM-135M"
-            });
+            let chunk = serde_json::json!({ "type": "llm_chunk", "token": text, "prompt": prompt, "model": "SmolLM-135M" });
             let _ = ws.borrow().send_with_str(&chunk.to_string());
         }
+        tokens_generated += 1;
+    }
+
+    // Autoregressiivinen generointi: yksi token kerrallaan
+    for _ in 1..max_new_tokens {
+        if next_token == 2 { break; }
+
+        let input = match Tensor::new(&[next_token], &device).and_then(|t| t.unsqueeze(0)) {
+            Ok(t) => t,
+            Err(e) => { console_log!("[SmolLM] Tensor-virhe: {}", e); break; }
+        };
+
+        let logits = match model.forward(&input, pos, &mut cache) {
+            Ok(l) => l,
+            Err(e) => { console_log!("[SmolLM] Forward-virhe pos {}: {}", pos, e); break; }
+        };
+
+        let logits = logits.squeeze(0).unwrap();
+        let logits = if logits.dims().len() == 2 {
+            logits.get(logits.dim(0).unwrap() - 1).unwrap()
+        } else {
+            logits
+        };
+        next_token = logits.argmax(0).unwrap().to_vec0::<u32>().unwrap();
+        pos += 1;
+
+        if next_token == 2 { break; }
+
+        if let Ok(text) = tokenizer.decode(&[next_token], true) {
+            generated_text.push_str(&text);
+            let chunk = serde_json::json!({ "type": "llm_chunk", "token": text, "prompt": prompt, "model": "SmolLM-135M" });
+            let _ = ws.borrow().send_with_str(&chunk.to_string());
+        }
+        tokens_generated += 1;
     }
 
     let gen_time = perf.now() - start_gen;
-    let tokens_generated = all_tokens.len() - input_len;
     let tokens_per_sec = if gen_time > 0.0 { (tokens_generated as f64 / gen_time) * 1000.0 } else { 0.0 };
 
     console_log!("[SmolLM] Generoitu {} tokenia | {:.0}ms | {:.1} tok/s", tokens_generated, gen_time, tokens_per_sec);

network-poc/static/index.html

@@ -339,16 +339,16 @@
                     <label class="task-option" data-task="qwen-05b">
                         <input type="radio" name="task" value="qwen-05b" style="display:none">
                         <div class="task-title">Qwen2.5 0.5B</div>
-                        <div class="task-desc">Tehokkaampi kielimalli — vaatii WebGPU:n ja vähintään 1 GB muistia</div>
-                        <div class="task-size">Lataus: ~430 MB (Q4)</div>
-                        <span class="task-badge task-soon">Tulossa</span>
+                        <div class="task-desc">Tehokkaampi kielimalli — vaatii vähintään 2 GB muistia (CPU)</div>
+                        <div class="task-size">Lataus: ~990 MB (safetensors) + 7 MB (tokenizer)</div>
+                        <span class="task-badge task-ready">Valmis</span>
                     </label>
                     <label class="task-option" data-task="phi3-mini">
                         <input type="radio" name="task" value="phi3-mini" style="display:none">
                         <div class="task-title">Phi-3 Mini 3.8B</div>
-                        <div class="task-desc">Iso kielimalli — vaatii tehokkaan GPU:n ja vähintään 4 GB VRAM</div>
-                        <div class="task-size">Lataus: ~2.3 GB (Q4)</div>
-                        <span class="task-badge task-soon">Tulossa</span>
+                        <div class="task-desc">Iso kielimalli — vaatii native-noden (Docker + GPU)</div>
+                        <div class="task-size">~7.6 GB — liian suuri selaimelle</div>
+                        <span class="task-badge task-soon">Vain native</span>
                     </label>
                 </div>
             </div>