1use crate::model_config;
10use super::traits::*;
11use anyhow::Result;
12use serde::{Serialize, Deserialize};
13
14model_config!(OlmoConfig {
16 vocab_size: usize = 50304,
17 hidden_size: usize = 4096,
18 intermediate_size: usize = 11008,
19 num_hidden_layers: usize = 32,
20 num_attention_heads: usize = 32,
21 num_key_value_heads: usize = 32,
22 hidden_act: String = "silu".to_string(),
23 max_position_embeddings: usize = 2048,
24 initializer_range: f32 = 0.02,
25 layer_norm_eps: f32 = 1e-5,
26 use_cache: bool = true,
27 pad_token_id: i64 = 1,
28 bos_token_id: i64 = 0,
29 eos_token_id: i64 = 0,
30 tie_word_embeddings: bool = false,
31 rope_theta: f32 = 10000.0,
32});
33
34impl OlmoConfig {
35 pub fn from_gguf_config(gguf: &crate::weight_loader_core::GGUFModelConfig) -> Self {
36 Self {
37 vocab_size: gguf.vocab_size,
38 hidden_size: gguf.hidden_size,
39 intermediate_size: gguf.intermediate_size,
40 num_hidden_layers: gguf.num_hidden_layers,
41 num_attention_heads: gguf.num_attention_heads,
42 num_key_value_heads: gguf.num_key_value_heads,
43 max_position_embeddings: gguf.max_position_embeddings,
44 rope_theta: gguf.rope_theta,
45 ..Default::default()
46 }
47 }
48}
49
50pub struct OlmoModelV2 {
51 config: OlmoConfig,
52 device: Device,
53 embed_tokens: Tensor,
54 layers: Vec<OlmoLayer>,
55 norm: Tensor,
56 lm_head: Tensor,
57}
58
59pub struct OlmoLayer {
60 self_attn: OlmoAttention,
61 mlp: OlmoMLP,
62 input_layernorm: Tensor,
63 post_attention_layernorm: Tensor,
64}
65
66pub struct OlmoAttention {
67 q_proj: Tensor,
68 k_proj: Tensor,
69 v_proj: Tensor,
70 o_proj: Tensor,
71 num_heads: usize,
72 num_kv_heads: usize,
73 head_dim: usize,
74 scale: f32,
75}
76
77pub struct OlmoMLP {
78 gate_proj: Tensor,
79 up_proj: Tensor,
80 down_proj: Tensor,
81}
82
83fn apply_rope_olmo(
84 q: &candle_core::Tensor,
85 k: &candle_core::Tensor,
86 seq_len: usize,
87 head_dim: usize,
88 rope_theta: f32,
89) -> Result<(candle_core::Tensor, candle_core::Tensor)> {
90 let device = q.device();
91 let half_dim = head_dim / 2;
92 let inv_freq: Vec<f32> = (0..half_dim)
93 .map(|i| 1.0 / rope_theta.powf((2 * i) as f32 / head_dim as f32))
94 .collect();
95
96 let positions: Vec<f32> = (0..seq_len).map(|p| p as f32).collect();
97 let mut angles = Vec::with_capacity(seq_len * half_dim);
98 for pos in &positions {
99 for freq in &inv_freq {
100 angles.push(pos * freq);
101 }
102 }
103
104 let angles_tensor = candle_core::Tensor::from_vec(angles, &[seq_len, half_dim], device)?;
105 let cos = angles_tensor.cos()?.unsqueeze(0)?.unsqueeze(0)?;
106 let sin = angles_tensor.sin()?.unsqueeze(0)?.unsqueeze(0)?;
107
108 let q_half1 = q.narrow(3, 0, half_dim)?;
109 let q_half2 = q.narrow(3, half_dim, half_dim)?;
110 let k_half1 = k.narrow(3, 0, half_dim)?;
111 let k_half2 = k.narrow(3, half_dim, half_dim)?;
112
113 let q_rot1 = (q_half1.broadcast_mul(&cos)? - q_half2.broadcast_mul(&sin)?)?;
114 let q_rot2 = (q_half1.broadcast_mul(&sin)? + q_half2.broadcast_mul(&cos)?)?;
115 let k_rot1 = (k_half1.broadcast_mul(&cos)? - k_half2.broadcast_mul(&sin)?)?;
116 let k_rot2 = (k_half1.broadcast_mul(&sin)? + k_half2.broadcast_mul(&cos)?)?;
117
118 Ok((
119 candle_core::Tensor::cat(&[&q_rot1, &q_rot2], 3)?,
120 candle_core::Tensor::cat(&[&k_rot1, &k_rot2], 3)?
121 ))
122}
123
124impl Model for OlmoModelV2 {
125 type Config = OlmoConfig;
126
127 fn new(config: OlmoConfig) -> Result<Self> {
128 let device = Device::CPU;
129 let embed_tokens = ops_fn::zeros(&[config.vocab_size, config.hidden_size], DataType::Float32, &device)?;
130 let norm = ops_fn::zeros(&[config.hidden_size], DataType::Float32, &device)?;
131 let lm_head = ops_fn::zeros(&[config.vocab_size, config.hidden_size], DataType::Float32, &device)?;
132
133 let mut layers = Vec::with_capacity(config.num_hidden_layers);
134 for _ in 0..config.num_hidden_layers {
135 layers.push(OlmoLayer::new(&config, &device)?);
136 }
137
138 Ok(Self { config, device, embed_tokens, layers, norm, lm_head })
139 }
140
141 fn from_weights(config: OlmoConfig, weights: ModelWeights) -> Result<Self> {
142 let mut model = Self::new(config)?;
143 if let Some(w) = weights.get("model.embed_tokens.weight") { model.embed_tokens = w.clone(); }
144 if let Some(w) = weights.get("model.norm.weight") { model.norm = w.clone(); }
145 if model.config.tie_word_embeddings {
146 model.lm_head = model.embed_tokens.clone();
147 } else if let Some(w) = weights.get("lm_head.weight") {
148 model.lm_head = w.clone();
149 }
150 for (i, layer) in model.layers.iter_mut().enumerate() { layer.load_weights(&weights, i)?; }
151 Ok(model)
152 }
153
154 fn forward(&self, inputs: &ModelInputs) -> Result<ModelOutputs> {
155 match inputs {
156 ModelInputs::Text { input_ids, .. } => {
157 let seq_len = input_ids.shape()[1];
158 let mut hidden = ops_fn::embedding(input_ids, &self.embed_tokens)?;
159
160 for layer in &self.layers {
161 hidden = layer.forward(&hidden, seq_len, self.config.rope_theta)?;
162 }
163
164 hidden = ops_fn::layer_norm(&hidden, &self.norm, None, self.config.layer_norm_eps)?;
165 let logits = ops_fn::matmul(&hidden, &ops_fn::transpose(&self.lm_head)?)?;
166
167 Ok(ModelOutputs::Logits { logits, hidden_states: None })
168 }
169 _ => Err(anyhow::anyhow!("OLMo only supports text inputs")),
170 }
171 }
172
173 fn generate(&self, prompt: &str, config: &GenerationConfig) -> Result<String> {
174 use crate::tokenizer::Tokenizer;
175 use rand::Rng;
176 let tokenizer = Tokenizer::new();
177 let mut tokens: Vec<u32> = tokenizer.encode(prompt);
178 for _ in 0..config.max_new_tokens {
179 let tokens_i64: Vec<i64> = tokens.iter().map(|&t| t as i64).collect();
180 let input = Tensor::from_i64_slice(&tokens_i64, &[1, tokens.len()], &self.device)?;
181 let outputs = self.forward(&ModelInputs::text(input))?;
182 let logits = match outputs { ModelOutputs::Logits { logits, .. } => logits, _ => return Err(anyhow::anyhow!("Expected logits")) };
183 let logits_candle = logits.to_candle()?;
184 let last = logits_candle.narrow(1, logits_candle.dims()[1] - 1, 1)?.squeeze(1)?.squeeze(0)?;
185 let logits_vec: Vec<f32> = last.to_vec1()?;
186 let next = if config.do_sample && config.temperature > 0.0 {
187 let scaled: Vec<f32> = logits_vec.iter().map(|&x| x / config.temperature).collect();
188 let max_v = scaled.iter().cloned().fold(f32::NEG_INFINITY, f32::max);
189 let exp_sum: f32 = scaled.iter().map(|&x| (x - max_v).exp()).sum();
190 let probs: Vec<f32> = scaled.iter().map(|&x| (x - max_v).exp() / exp_sum).collect();
191 let mut rng = rand::thread_rng();
192 let r: f32 = rng.gen();
193 let mut cum = 0.0;
194 let mut s = 0u32;
195 for (i, &p) in probs.iter().enumerate() { cum += p; if r <= cum { s = i as u32; break; } }
196 s
197 } else {
198 logits_vec.iter().enumerate().max_by(|a, b| a.1.partial_cmp(b.1).unwrap()).map(|(i, _)| i as u32).unwrap_or(0)
199 };
200 if next == config.eos_token_id { break; }
201 tokens.push(next);
202 }
203 Ok(tokenizer.decode(&tokens))
204 }
205
206 fn config(&self) -> &Self::Config { &self.config }
207 fn memory_requirements(&self) -> MemoryRequirements {
208 let p = self.config.vocab_size * self.config.hidden_size + self.config.num_hidden_layers * 8 * self.config.hidden_size.pow(2);
209 MemoryRequirements { gpu_memory: p * 4, cpu_memory: p, kv_cache_memory: 2 * self.config.num_hidden_layers * self.config.max_position_embeddings * self.config.hidden_size * 4, peak_memory: p * 5 }
210 }
211 fn to_device(&mut self, device: &Device) -> Result<()> {
212 self.embed_tokens = self.embed_tokens.to_device(device)?;
213 self.norm = self.norm.to_device(device)?;
214 self.lm_head = self.lm_head.to_device(device)?;
215 for layer in &mut self.layers { layer.to_device(device)?; }
216 self.device = device.clone();
217 Ok(())
218 }
219}
220
221impl OlmoLayer {
222 fn new(config: &OlmoConfig, device: &Device) -> Result<Self> {
223 Ok(Self {
224 self_attn: OlmoAttention::new(config, device)?,
225 mlp: OlmoMLP::new(config, device)?,
226 input_layernorm: ops_fn::zeros(&[config.hidden_size], DataType::Float32, device)?,
227 post_attention_layernorm: ops_fn::zeros(&[config.hidden_size], DataType::Float32, device)?,
228 })
229 }
230
231 fn forward(&self, hidden_states: &Tensor, seq_len: usize, rope_theta: f32) -> Result<Tensor> {
232 let residual = hidden_states.clone();
233 let h = ops_fn::layer_norm(hidden_states, &self.input_layernorm, None, 1e-5)?;
234 let attn_out = self.self_attn.forward(&h, seq_len, rope_theta)?;
235 let h = ops_fn::add(&residual, &attn_out)?;
236
237 let residual = h.clone();
238 let h = ops_fn::layer_norm(&h, &self.post_attention_layernorm, None, 1e-5)?;
239 let mlp_out = self.mlp.forward(&h)?;
240 ops_fn::add(&residual, &mlp_out)
241 }
242
243 fn load_weights(&mut self, weights: &ModelWeights, idx: usize) -> Result<()> {
244 let p = format!("model.layers.{}", idx);
245 if let Some(w) = weights.get(&format!("{}.self_attn.q_proj.weight", p)) { self.self_attn.q_proj = ops_fn::transpose(w)?; }
246 if let Some(w) = weights.get(&format!("{}.self_attn.k_proj.weight", p)) { self.self_attn.k_proj = ops_fn::transpose(w)?; }
247 if let Some(w) = weights.get(&format!("{}.self_attn.v_proj.weight", p)) { self.self_attn.v_proj = ops_fn::transpose(w)?; }
248 if let Some(w) = weights.get(&format!("{}.self_attn.o_proj.weight", p)) { self.self_attn.o_proj = ops_fn::transpose(w)?; }
249 if let Some(w) = weights.get(&format!("{}.mlp.gate_proj.weight", p)) { self.mlp.gate_proj = ops_fn::transpose(w)?; }
250 if let Some(w) = weights.get(&format!("{}.mlp.up_proj.weight", p)) { self.mlp.up_proj = ops_fn::transpose(w)?; }
251 if let Some(w) = weights.get(&format!("{}.mlp.down_proj.weight", p)) { self.mlp.down_proj = ops_fn::transpose(w)?; }
252 if let Some(w) = weights.get(&format!("{}.input_layernorm.weight", p)) { self.input_layernorm = w.clone(); }
253 if let Some(w) = weights.get(&format!("{}.post_attention_layernorm.weight", p)) { self.post_attention_layernorm = w.clone(); }
254 Ok(())
255 }
256
257 fn to_device(&mut self, device: &Device) -> Result<()> {
258 self.self_attn.to_device(device)?;
259 self.mlp.to_device(device)?;
260 self.input_layernorm = self.input_layernorm.to_device(device)?;
261 self.post_attention_layernorm = self.post_attention_layernorm.to_device(device)?;
262 Ok(())
263 }
264}
265
266impl OlmoAttention {
267 fn new(config: &OlmoConfig, device: &Device) -> Result<Self> {
268 let head_dim = config.hidden_size / config.num_attention_heads;
269 Ok(Self {
270 q_proj: ops_fn::zeros(&[config.hidden_size, config.num_attention_heads * head_dim], DataType::Float32, device)?,
271 k_proj: ops_fn::zeros(&[config.hidden_size, config.num_key_value_heads * head_dim], DataType::Float32, device)?,
272 v_proj: ops_fn::zeros(&[config.hidden_size, config.num_key_value_heads * head_dim], DataType::Float32, device)?,
273 o_proj: ops_fn::zeros(&[config.num_attention_heads * head_dim, config.hidden_size], DataType::Float32, device)?,
274 num_heads: config.num_attention_heads,
275 num_kv_heads: config.num_key_value_heads,
276 head_dim,
277 scale: 1.0 / (head_dim as f32).sqrt(),
278 })
279 }
280
281 fn forward(&self, hidden_states: &Tensor, seq_len: usize, rope_theta: f32) -> Result<Tensor> {
282 let shape = hidden_states.shape();
283 let batch = shape[0];
284
285 let q = ops_fn::matmul(hidden_states, &self.q_proj)?.to_candle()?;
286 let k = ops_fn::matmul(hidden_states, &self.k_proj)?.to_candle()?;
287 let v = ops_fn::matmul(hidden_states, &self.v_proj)?.to_candle()?;
288
289 let q = q.reshape(&[batch, seq_len, self.num_heads, self.head_dim])?.transpose(1, 2)?;
290 let k = k.reshape(&[batch, seq_len, self.num_kv_heads, self.head_dim])?.transpose(1, 2)?;
291 let v = v.reshape(&[batch, seq_len, self.num_kv_heads, self.head_dim])?.transpose(1, 2)?;
292
293 let (q, k) = apply_rope_olmo(&q, &k, seq_len, self.head_dim, rope_theta)?;
294
295 let num_groups = self.num_heads / self.num_kv_heads;
297 let (k, v) = if num_groups > 1 {
298 let k_exp = k.unsqueeze(2)?.broadcast_as(&[batch, self.num_kv_heads, num_groups, seq_len, self.head_dim])?.reshape(&[batch, self.num_heads, seq_len, self.head_dim])?;
299 let v_exp = v.unsqueeze(2)?.broadcast_as(&[batch, self.num_kv_heads, num_groups, seq_len, self.head_dim])?.reshape(&[batch, self.num_heads, seq_len, self.head_dim])?;
300 (k_exp, v_exp)
301 } else {
302 (k, v)
303 };
304
305 let q = q.contiguous()?;
306 let k_t = k.transpose(2, 3)?.contiguous()?;
307 let scores = (q.matmul(&k_t)? * (self.scale as f64))?;
308
309 let device = scores.device();
310 let mask = {
311 let mut m = vec![0.0f32; seq_len * seq_len];
312 for i in 0..seq_len { for j in (i+1)..seq_len { m[i*seq_len+j] = f32::NEG_INFINITY; } }
313 candle_core::Tensor::from_vec(m, &[1, 1, seq_len, seq_len], device)?
314 };
315 let scores = scores.broadcast_add(&mask)?;
316
317 let v = v.contiguous()?;
318 let attn = candle_nn::ops::softmax_last_dim(&scores)?.matmul(&v)?;
319 let out = attn.transpose(1, 2)?.reshape(&[batch, seq_len, self.num_heads * self.head_dim])?;
320 ops_fn::matmul(&Tensor::from_candle(out), &self.o_proj)
321 }
322
323 fn to_device(&mut self, device: &Device) -> Result<()> {
324 self.q_proj = self.q_proj.to_device(device)?;
325 self.k_proj = self.k_proj.to_device(device)?;
326 self.v_proj = self.v_proj.to_device(device)?;
327 self.o_proj = self.o_proj.to_device(device)?;
328 Ok(())
329 }
330}
331
332impl OlmoMLP {
333 fn new(config: &OlmoConfig, device: &Device) -> Result<Self> {
334 Ok(Self {
335 gate_proj: ops_fn::zeros(&[config.hidden_size, config.intermediate_size], DataType::Float32, device)?,
336 up_proj: ops_fn::zeros(&[config.hidden_size, config.intermediate_size], DataType::Float32, device)?,
337 down_proj: ops_fn::zeros(&[config.intermediate_size, config.hidden_size], DataType::Float32, device)?,
338 })
339 }
340 fn forward(&self, x: &Tensor) -> Result<Tensor> {
341 let gate = ops_fn::matmul(x, &self.gate_proj)?;
342 let up = ops_fn::matmul(x, &self.up_proj)?;
343 let h = ops_fn::mul(&ops_fn::silu(&gate)?, &up)?;
344 ops_fn::matmul(&h, &self.down_proj)
345 }
346 fn to_device(&mut self, device: &Device) -> Result<()> {
347 self.gate_proj = self.gate_proj.to_device(device)?;
348 self.up_proj = self.up_proj.to_device(device)?;
349 self.down_proj = self.down_proj.to_device(device)?;
350 Ok(())
351 }
352}
353
354#[cfg(test)]
355mod tests {
356 use super::*;
357 #[test]
358 fn test_olmo_creation() {
359 let config = OlmoConfig { vocab_size: 1000, hidden_size: 128, intermediate_size: 512, num_hidden_layers: 2, num_attention_heads: 4, num_key_value_heads: 4, ..Default::default() };
360 let model = OlmoModelV2::new(config).unwrap();
361 assert_eq!(model.config().vocab_size(), 1000);
362 }
363 #[test]
364 fn test_olmo_forward() {
365 let config = OlmoConfig { vocab_size: 100, hidden_size: 64, intermediate_size: 256, num_hidden_layers: 1, num_attention_heads: 4, num_key_value_heads: 4, ..Default::default() };
366 let model = OlmoModelV2::new(config).unwrap();
367 let inputs = ModelInputs::text(ops_fn::zeros(&[2, 8], DataType::Int64, &Device::CPU).unwrap());
368 match model.forward(&inputs).unwrap() { ModelOutputs::Logits { logits, .. } => assert_eq!(logits.shape(), &[2, 8, 100]), _ => panic!() }
369 }
370}