kylä lähtee!

network-poc/node/src/burn_smollm/attention.rs

@@ -0,0 +1,118 @@
+use burn::module::{Module, Param};
+use burn::tensor::{backend::Backend, Tensor};
+use super::rope::RoPE;
+use super::config::SmolLMConfig;
+
+#[derive(Clone, Debug)]
+pub struct KVCache<B: Backend> {
+    pub k: Tensor<B, 4>,
+    pub v: Tensor<B, 4>,
+}
+
+#[derive(Module, Debug)]
+pub struct Attention<B: Backend> {
+    pub q_proj: Param<Tensor<B, 2>>,  // [hidden, num_heads * head_dim]
+    pub k_proj: Param<Tensor<B, 2>>,  // [hidden, num_kv_heads * head_dim]
+    pub v_proj: Param<Tensor<B, 2>>,  // [hidden, num_kv_heads * head_dim]
+    pub o_proj: Param<Tensor<B, 2>>,  // [num_heads * head_dim, hidden]
+
+    num_heads: usize,
+    num_kv_heads: usize,
+    head_dim: usize,
+
+    rope: RoPE<B>,
+}
+
+impl<B: Backend> Attention<B> {
+    pub fn new(config: &SmolLMConfig, device: &B::Device) -> Self {
+        let head_dim = config.hidden_size / config.num_attention_heads;
+
+        Self {
+            q_proj: Param::from_tensor(Tensor::zeros([config.hidden_size, config.num_attention_heads * head_dim], device)),
+            k_proj: Param::from_tensor(Tensor::zeros([config.hidden_size, config.num_key_value_heads * head_dim], device)),
+            v_proj: Param::from_tensor(Tensor::zeros([config.hidden_size, config.num_key_value_heads * head_dim], device)),
+            o_proj: Param::from_tensor(Tensor::zeros([config.num_attention_heads * head_dim, config.hidden_size], device)),
+
+            num_heads: config.num_attention_heads,
+            num_kv_heads: config.num_key_value_heads,
+            head_dim,
+
+            rope: RoPE::new(head_dim, config.max_position_embeddings, config.rope_theta, device),
+        }
+    }
+
+    pub fn forward(
+        &self,
+        x: Tensor<B, 3>,
+        offset: usize,
+        cache: Option<KVCache<B>>
+    ) -> (Tensor<B, 3>, KVCache<B>) {
+        let [batch, seq_len, hidden_dim] = x.dims();
+
+        // Project Q, K, V: x @ W -> [batch, seq, proj_dim]
+        let q = x.clone().matmul(self.q_proj.val().unsqueeze());
+        let k = x.clone().matmul(self.k_proj.val().unsqueeze());
+        let v = x.matmul(self.v_proj.val().unsqueeze());
+
+        // Reshape: [batch, seq, heads, head_dim] -> [batch, heads, seq, head_dim]
+        let q = q.reshape([batch, seq_len, self.num_heads, self.head_dim]).swap_dims(1, 2);
+        let k = k.reshape([batch, seq_len, self.num_kv_heads, self.head_dim]).swap_dims(1, 2);
+        let v = v.reshape([batch, seq_len, self.num_kv_heads, self.head_dim]).swap_dims(1, 2);
+
+        // Apply RoPE
+        let q = self.rope.forward(q, offset);
+        let k = self.rope.forward(k, offset);
+
+        // KV cache
+        let (k, v) = if let Some(c) = cache {
+            (Tensor::cat(vec![c.k, k], 2), Tensor::cat(vec![c.v, v], 2))
+        } else {
+            (k, v)
+        };
+
+        let new_cache = KVCache { k: k.clone(), v: v.clone() };
+        let kv_len = k.dims()[2];
+
+        // GQA: repeat K,V heads — [batch, kv_heads, kv_len, hd] -> [batch, num_heads, kv_len, hd]
+        let num_reps = self.num_heads / self.num_kv_heads;
+        let k = if num_reps > 1 {
+            let [b, kv_h, s, hd] = k.dims();
+            k.reshape([b, kv_h, 1, s, hd]).repeat_dim(2, num_reps).reshape([b, self.num_heads, s, hd])
+        } else { k };
+        let v = if num_reps > 1 {
+            let [b, kv_h, s, hd] = v.dims();
+            v.reshape([b, kv_h, 1, s, hd]).repeat_dim(2, num_reps).reshape([b, self.num_heads, s, hd])
+        } else { v };
+
+        // Attention: Q @ K^T / sqrt(d)
+        let scale = 1.0 / (self.head_dim as f64).sqrt();
+        let scores = q.matmul(k.swap_dims(2, 3)).mul_scalar(scale);
+        // scores: [batch, heads, seq_len, kv_len]
+
+        // Causal mask for prefill (seq_len > 1)
+        let scores = if seq_len > 1 {
+            let mask_data: Vec<f32> = (0..seq_len).flat_map(|i| {
+                (0..kv_len).map(move |j| {
+                    if j > offset + i { f32::NEG_INFINITY } else { 0.0 }
+                })
+            }).collect();
+            let mask = Tensor::<B, 2>::from_data(
+                burn::tensor::TensorData::new(mask_data, [seq_len, kv_len]),
+                &scores.device()
+            ).reshape([1, 1, seq_len, kv_len]);
+            scores + mask
+        } else {
+            scores
+        };
+
+        let attn_weights = burn::tensor::activation::softmax(scores, 3);
+
+        let context = attn_weights.matmul(v);
+        // [batch, heads, seq, hd] -> [batch, seq, heads*hd]
+        let context = context.swap_dims(1, 2).reshape([batch, seq_len, self.num_heads * self.head_dim]);
+
+        let output = context.matmul(self.o_proj.val().unsqueeze());
+
+        (output, new_cache)
+    }
+}

network-poc/node/src/burn_smollm/config.rs

@@ -0,0 +1,28 @@
+#[derive(Clone, Debug)]
+pub struct SmolLMConfig {
+    pub hidden_size: usize,
+    pub intermediate_size: usize,
+    pub vocab_size: usize,
+    pub num_hidden_layers: usize,
+    pub num_attention_heads: usize,
+    pub num_key_value_heads: usize,
+    pub rms_norm_eps: f64,
+    pub rope_theta: f32,
+    pub max_position_embeddings: usize,
+}
+
+impl Default for SmolLMConfig {
+    fn default() -> Self {
+        Self {
+            hidden_size: 576,
+            intermediate_size: 1536,
+            vocab_size: 49152,
+            num_hidden_layers: 30,
+            num_attention_heads: 9,
+            num_key_value_heads: 3,
+            rms_norm_eps: 1e-5,
+            rope_theta: 10000.0,
+            max_position_embeddings: 2048,
+        }
+    }
+}

network-poc/node/src/burn_smollm/loader.rs

@@ -0,0 +1,90 @@
+use burn::tensor::{backend::Backend, Tensor, Data};
+use candle_core::safetensors;
+use candle_core::Device as CandleDevice;
+use burn::module::Param;
+use super::model::LlamaModel;
+use super::config::SmolLMConfig;
+
+fn load_tensor_2d<B: Backend>(
+    tensors_map: &std::collections::HashMap<String, candle_core::Tensor>,
+    name: &str,
+    device: &B::Device,
+    shape_out_in: [usize; 2]
+) -> Result<Param<Tensor<B, 2>>, String> {
+    let t = tensors_map.get(name).ok_or_else(|| format!("Puuttuu: {}", name))?;
+    let t = t.to_dtype(candle_core::DType::F32).unwrap();
+    let vec = t.flatten_all().unwrap().to_vec1::<f32>().unwrap();
+    let t_burn = Tensor::<B, 2>::from_data(burn::tensor::TensorData::new(vec, shape_out_in), device);
+    // transpose from [out, in] to [in, out]
+    Ok(Param::from_tensor(t_burn.transpose()))
+}
+
+fn load_tensor_1d<B: Backend>(
+    tensors_map: &std::collections::HashMap<String, candle_core::Tensor>,
+    name: &str,
+    device: &B::Device,
+    _shape: [usize; 1]
+) -> Result<Param<Tensor<B, 1>>, String> {
+    let t = tensors_map.get(name).ok_or_else(|| format!("Puuttuu: {}", name))?;
+    let t = t.to_dtype(candle_core::DType::F32).unwrap();
+    let vec = t.flatten_all().unwrap().to_vec1::<f32>().unwrap();
+    Ok(Param::from_tensor(Tensor::<B, 1>::from_floats(vec.as_slice(), device)))
+}
+
+fn load_embed<B: Backend>(
+    tensors_map: &std::collections::HashMap<String, candle_core::Tensor>,
+    name: &str,
+    device: &B::Device,
+    shape: [usize; 2]
+) -> Result<Param<Tensor<B, 2>>, String> {
+    let t = tensors_map.get(name).ok_or_else(|| format!("Puuttuu: {}", name))?;
+    let t = t.to_dtype(candle_core::DType::F32).unwrap();
+    let vec = t.flatten_all().unwrap().to_vec1::<f32>().unwrap();
+    // Embed ei transponoi samalla tavalla, se pysyy [vocab, hidden]
+    Ok(Param::from_tensor(Tensor::<B, 2>::from_data(burn::tensor::TensorData::new(vec, shape), device)))
+}
+
+pub fn load_safetensors_to_model<B: Backend>(
+    buffer: &[u8], 
+    config: &SmolLMConfig, 
+    device: &B::Device
+) -> Result<LlamaModel<B>, String> {
+    
+    let mut model = LlamaModel::new(config, device);
+    let tensors_map = safetensors::load_buffer(buffer, &CandleDevice::Cpu)
+        .map_err(|e| format!("Virhe Safetensors luennassa: {}", e))?;
+    
+    // Embeddings
+    model.embed_tokens = load_embed(&tensors_map, "model.embed_tokens.weight", device, [config.vocab_size, config.hidden_size])?;
+    model.norm.weight = load_tensor_1d(&tensors_map, "model.norm.weight", device, [config.hidden_size])?;
+    model.lm_head = load_embed(&tensors_map, "lm_head.weight", device, [config.vocab_size, config.hidden_size]).or_else(|_| {
+        load_embed(&tensors_map, "model.embed_tokens.weight", device, [config.vocab_size, config.hidden_size])
+    })?;
+    
+    let head_dim = config.hidden_size / config.num_attention_heads;
+
+    for i in 0..config.num_hidden_layers {
+        let prefix = format!("model.layers.{}", i);
+        
+        let layer = &mut model.layers[i];
+        
+        // Norms
+        layer.input_layernorm.weight = load_tensor_1d(&tensors_map, &format!("{}.input_layernorm.weight", prefix), device, [config.hidden_size])?;
+        layer.post_attention_layernorm.weight = load_tensor_1d(&tensors_map, &format!("{}.post_attention_layernorm.weight", prefix), device, [config.hidden_size])?;
+        
+        // Attention
+        let num_heads = config.num_attention_heads;
+        let num_kv_heads = config.num_key_value_heads;
+        layer.self_attn.q_proj = load_tensor_2d(&tensors_map, &format!("{}.self_attn.q_proj.weight", prefix), device, [num_heads * head_dim, config.hidden_size])?;
+        layer.self_attn.k_proj = load_tensor_2d(&tensors_map, &format!("{}.self_attn.k_proj.weight", prefix), device, [num_kv_heads * head_dim, config.hidden_size])?;
+        layer.self_attn.v_proj = load_tensor_2d(&tensors_map, &format!("{}.self_attn.v_proj.weight", prefix), device, [num_kv_heads * head_dim, config.hidden_size])?;
+        layer.self_attn.o_proj = load_tensor_2d(&tensors_map, &format!("{}.self_attn.o_proj.weight", prefix), device, [config.hidden_size, num_heads * head_dim])?;
+        
+        // MLP
+        layer.mlp.gate_proj = load_tensor_2d(&tensors_map, &format!("{}.mlp.gate_proj.weight", prefix), device, [config.intermediate_size, config.hidden_size])?;
+        layer.mlp.up_proj = load_tensor_2d(&tensors_map, &format!("{}.mlp.up_proj.weight", prefix), device, [config.intermediate_size, config.hidden_size])?;
+        layer.mlp.down_proj = load_tensor_2d(&tensors_map, &format!("{}.mlp.down_proj.weight", prefix), device, [config.hidden_size, config.intermediate_size])?;
+    }
+    
+    Ok(model)
+}

network-poc/node/src/burn_smollm/mod.rs

@@ -0,0 +1,6 @@
+pub mod attention;
+pub mod config;
+pub mod loader;
+pub mod model;
+pub mod modules;
+pub mod rope;

network-poc/node/src/burn_smollm/model.rs

@@ -0,0 +1,96 @@
+use burn::module::{Module, Param};
+use burn::tensor::{backend::Backend, Tensor, Int};
+use super::modules::{RmsNorm, Mlp};
+use super::attention::{Attention, KVCache};
+use super::config::SmolLMConfig;
+
+#[derive(Module, Debug)]
+pub struct LlamaBlock<B: Backend> {
+    pub self_attn: Attention<B>,
+    pub mlp: Mlp<B>,
+    pub input_layernorm: RmsNorm<B>,
+    pub post_attention_layernorm: RmsNorm<B>,
+}
+
+impl<B: Backend> LlamaBlock<B> {
+    pub fn new(config: &SmolLMConfig, device: &B::Device) -> Self {
+        Self {
+            self_attn: Attention::new(config, device),
+            mlp: Mlp::new(config.hidden_size, config.intermediate_size, device),
+            input_layernorm: RmsNorm::new(config.hidden_size, config.rms_norm_eps, device),
+            post_attention_layernorm: RmsNorm::new(config.hidden_size, config.rms_norm_eps, device),
+        }
+    }
+
+    pub fn forward(
+        &self,
+        x: Tensor<B, 3>,
+        offset: usize,
+        cache: Option<KVCache<B>>
+    ) -> (Tensor<B, 3>, KVCache<B>) {
+        let residual = x.clone();
+        let x_norm = self.input_layernorm.forward(x);
+        
+        let (attn_out, new_cache) = self.self_attn.forward(x_norm, offset, cache);
+        
+        let x = residual + attn_out;
+        
+        let residual = x.clone();
+        let x_norm = self.post_attention_layernorm.forward(x);
+        let mlp_out = self.mlp.forward(x_norm);
+        
+        let x = residual + mlp_out;
+        (x, new_cache)
+    }
+}
+
+#[derive(Module, Debug)]
+pub struct LlamaModel<B: Backend> {
+    pub embed_tokens: Param<Tensor<B, 2>>,
+    pub layers: Vec<LlamaBlock<B>>,
+    pub norm: RmsNorm<B>,
+    pub lm_head: Param<Tensor<B, 2>>, // For tie_word_embeddings this can point to embed_tokens
+}
+
+impl<B: Backend> LlamaModel<B> {
+    pub fn new(config: &SmolLMConfig, device: &B::Device) -> Self {
+        let embed = Tensor::zeros([config.vocab_size, config.hidden_size], device);
+        let lm_head = Tensor::zeros([config.vocab_size, config.hidden_size], device);
+        
+        let mut layers = Vec::new();
+        for _ in 0..config.num_hidden_layers {
+            layers.push(LlamaBlock::new(config, device));
+        }
+
+        Self {
+            embed_tokens: Param::from_tensor(embed),
+            layers,
+            norm: RmsNorm::new(config.hidden_size, config.rms_norm_eps, device),
+            lm_head: Param::from_tensor(lm_head),
+        }
+    }
+
+    pub fn forward(
+        &self,
+        input_ids: Tensor<B, 2, Int>,
+        offset: usize,
+        caches: &mut Vec<Option<KVCache<B>>>
+    ) -> Tensor<B, 3> {
+        let [_batch, _seq_len] = input_ids.dims();
+        
+        let mut x = burn::tensor::module::embedding(self.embed_tokens.val(), input_ids);
+        
+        for (i, layer) in self.layers.iter().enumerate() {
+            let cache = caches[i].take();
+            let (out, new_cache) = layer.forward(x, offset, cache);
+            x = out;
+            caches[i] = Some(new_cache);
+        }
+        
+        x = self.norm.forward(x);
+        
+        // Matmul with lm_head (or embed_tokens if tied) to get logits
+        // Notice: lm_head is typically [vocab_size, hidden_size] in HF, so we swap dims
+        x.matmul(self.lm_head.val().swap_dims(0, 1).unsqueeze())
+    }
+}

network-poc/node/src/burn_smollm/modules.rs

@@ -0,0 +1,59 @@
+use burn::module::{Module, Param};
+use burn::tensor::{backend::Backend, Tensor};
+
+#[derive(Module, Debug)]
+pub struct RmsNorm<B: Backend> {
+    pub weight: Param<Tensor<B, 1>>,
+    epsilon: f64,
+}
+
+impl<B: Backend> RmsNorm<B> {
+    pub fn new(size: usize, epsilon: f64, device: &B::Device) -> Self {
+        let weight = Param::from_tensor(Tensor::ones([size], device));
+        Self { weight, epsilon }
+    }
+
+    pub fn forward(&self, x: Tensor<B, 3>) -> Tensor<B, 3> {
+        // x: [batch, seq_len, dim]
+        // RMSNorm: x * weight / sqrt(mean(x^2) + eps)
+        let x_sq = x.clone().powf_scalar(2.0);
+        // mean over last dim, keeping dims for broadcast
+        let [b, s, d] = x_sq.dims();
+        let variance = x_sq.sum_dim(2).div_scalar(d as f32);
+        let norm = x.div(variance.add_scalar(self.epsilon).sqrt());
+
+        let w = self.weight.val().unsqueeze::<2>().unsqueeze::<3>().reshape([1, 1, d]);
+        norm * w
+    }
+}
+
+#[derive(Module, Debug)]
+pub struct Mlp<B: Backend> {
+    pub gate_proj: Param<Tensor<B, 2>>, // [in, intermediate]
+    pub up_proj: Param<Tensor<B, 2>>,   // [in, intermediate]
+    pub down_proj: Param<Tensor<B, 2>>, // [intermediate, out]
+}
+
+impl<B: Backend> Mlp<B> {
+    pub fn new(hidden_size: usize, intermediate_size: usize, device: &B::Device) -> Self {
+        Self {
+            gate_proj: Param::from_tensor(Tensor::zeros([hidden_size, intermediate_size], device)),
+            up_proj: Param::from_tensor(Tensor::zeros([hidden_size, intermediate_size], device)),
+            down_proj: Param::from_tensor(Tensor::zeros([intermediate_size, hidden_size], device)),
+        }
+    }
+
+    pub fn forward(&self, x: Tensor<B, 3>) -> Tensor<B, 3> {
+        // x: [batch, seq, hidden]
+        // gate = x @ gate_proj  -> [batch, seq, intermediate]
+        let gate = x.clone().matmul(self.gate_proj.val().unsqueeze());
+        let up = x.matmul(self.up_proj.val().unsqueeze());
+
+        // SiLU(gate) * up
+        let silu = gate.clone() * burn::tensor::activation::sigmoid(gate);
+        let intermediate = silu * up;
+
+        // intermediate @ down_proj -> [batch, seq, hidden]
+        intermediate.matmul(self.down_proj.val().unsqueeze())
+    }
+}

network-poc/node/src/burn_smollm/rope.rs

@@ -0,0 +1,59 @@
+use burn::module::Module;
+use burn::tensor::{backend::Backend, Tensor};
+
+#[derive(Module, Debug)]
+pub struct RoPE<B: Backend> {
+    cos_cache: Tensor<B, 2>,
+    sin_cache: Tensor<B, 2>,
+}
+
+impl<B: Backend> RoPE<B> {
+    pub fn new(head_dim: usize, max_seq_len: usize, theta: f32, device: &B::Device) -> Self {
+        // (head_dim / 2) values
+        let half_dim = head_dim / 2;
+        let inv_freq: Vec<f32> = (0..half_dim)
+            .map(|i| 1.0 / theta.powf((2 * i) as f32 / head_dim as f32))
+            .collect();
+        
+        let inv_freq = Tensor::<B, 1>::from_floats(inv_freq.as_slice(), device).unsqueeze::<2>();
+        let t_floats: Vec<f32> = (0..max_seq_len).map(|v| v as f32).collect();
+        let t = Tensor::<B, 1>::from_floats(t_floats.as_slice(), device).unsqueeze::<2>().transpose();
+        // t shape: [max_seq_len, 1]
+        // inv_freq shape: [1, half_dim]
+        
+        // freqs shape: [max_seq_len, half_dim]
+        let freqs = t.matmul(inv_freq);
+        
+        let cos_cache = freqs.clone().cos();
+        let sin_cache = freqs.sin();
+
+        Self {
+            cos_cache,
+            sin_cache,
+        }
+    }
+
+    pub fn forward(&self, x: Tensor<B, 4>, offset: usize) -> Tensor<B, 4> {
+        let [batch, heads, seq_len, head_dim] = x.dims();
+        let half_dim = head_dim / 2;
+        
+        // x shape: [batch, heads, seq_len, head_dim]
+        // valitaan viipaleet (x1 ja x2) jotta saadaan pyöritettyä rotaatiot
+        let x1 = x.clone().slice([0..batch, 0..heads, 0..seq_len, 0..half_dim]);
+        let x2 = x.clone().slice([0..batch, 0..heads, 0..seq_len, half_dim..head_dim]);
+        
+        // haetaan vastaava seq offsetista alkaen
+        let cos = self.cos_cache.clone().slice([offset..offset+seq_len, 0..half_dim])
+            .unsqueeze::<4>() // [seq, half_dim, 1]
+            .reshape([1, 1, seq_len, half_dim]);
+        let sin = self.sin_cache.clone().slice([offset..offset+seq_len, 0..half_dim])
+            .reshape([1, 1, seq_len, half_dim]);
+
+        // x1 * cos - x2 * sin
+        let o1 = x1.clone().mul(cos.clone()) - x2.clone().mul(sin.clone());
+        // x2 * cos + x1 * sin
+        let o2 = x2.mul(cos) + x1.mul(sin);
+
+        Tensor::cat(vec![o1, o2], 3)
+    }
+}