1use crate::model_config;
9use super::traits::*;
10use anyhow::Result;
11use serde::{Serialize, Deserialize};
12
13model_config!(GrokConfig {
14 vocab_size: usize = 131072,
15 hidden_size: usize = 6144,
16 intermediate_size: usize = 32768,
17 num_hidden_layers: usize = 64,
18 num_attention_heads: usize = 48,
19 num_key_value_heads: usize = 8,
20 hidden_act: String = "gelu".to_string(),
21 max_position_embeddings: usize = 8192,
22 rms_norm_eps: f32 = 1e-5,
23 use_cache: bool = true,
24 pad_token_id: i64 = 0,
25 bos_token_id: i64 = 1,
26 eos_token_id: i64 = 2,
27 tie_word_embeddings: bool = false,
28 rope_theta: f32 = 10000.0,
29 num_experts: usize = 8,
30 num_experts_per_tok: usize = 2,
31});
32
33impl GrokConfig {
34 pub fn from_gguf_config(gguf: &crate::weight_loader_core::GGUFModelConfig) -> Self {
35 Self {
36 vocab_size: gguf.vocab_size,
37 hidden_size: gguf.hidden_size,
38 intermediate_size: gguf.intermediate_size,
39 num_hidden_layers: gguf.num_hidden_layers,
40 num_attention_heads: gguf.num_attention_heads,
41 num_key_value_heads: gguf.num_key_value_heads,
42 max_position_embeddings: gguf.max_position_embeddings,
43 rope_theta: gguf.rope_theta,
44 ..Default::default()
45 }
46 }
47}
48
49pub struct GrokModelV2 {
50 config: GrokConfig,
51 device: Device,
52 embed_tokens: Tensor,
53 layers: Vec<GrokLayer>,
54 norm: Tensor,
55 lm_head: Tensor,
56}
57
58pub struct GrokLayer {
59 self_attn: GrokAttention,
60 moe: GrokMoE,
61 input_layernorm: Tensor,
62 post_attention_layernorm: Tensor,
63}
64
65pub struct GrokAttention {
66 q_proj: Tensor,
67 k_proj: Tensor,
68 v_proj: Tensor,
69 o_proj: Tensor,
70 num_heads: usize,
71 num_kv_heads: usize,
72 head_dim: usize,
73 scale: f32,
74}
75
76pub struct GrokMoE {
77 router: Tensor,
78 experts: Vec<GrokExpert>,
79 num_experts_per_tok: usize,
80}
81
82pub struct GrokExpert {
83 gate_proj: Tensor,
84 up_proj: Tensor,
85 down_proj: Tensor,
86}
87
88fn apply_rope_grok(
89 q: &candle_core::Tensor, k: &candle_core::Tensor, seq_len: usize, head_dim: usize, rope_theta: f32,
90) -> Result<(candle_core::Tensor, candle_core::Tensor)> {
91 let device = q.device();
92 let half_dim = head_dim / 2;
93 let inv_freq: Vec<f32> = (0..half_dim).map(|i| 1.0 / rope_theta.powf((2 * i) as f32 / head_dim as f32)).collect();
94 let positions: Vec<f32> = (0..seq_len).map(|p| p as f32).collect();
95 let mut angles = Vec::with_capacity(seq_len * half_dim);
96 for pos in &positions { for freq in &inv_freq { angles.push(pos * freq); } }
97 let angles_tensor = candle_core::Tensor::from_vec(angles, &[seq_len, half_dim], device)?;
98 let cos = angles_tensor.cos()?.unsqueeze(0)?.unsqueeze(0)?;
99 let sin = angles_tensor.sin()?.unsqueeze(0)?.unsqueeze(0)?;
100 let (q_half1, q_half2) = (q.narrow(3, 0, half_dim)?, q.narrow(3, half_dim, half_dim)?);
101 let (k_half1, k_half2) = (k.narrow(3, 0, half_dim)?, k.narrow(3, half_dim, half_dim)?);
102 let q_rot = candle_core::Tensor::cat(&[&(q_half1.broadcast_mul(&cos)? - q_half2.broadcast_mul(&sin)?)?, &(q_half1.broadcast_mul(&sin)? + q_half2.broadcast_mul(&cos)?)?], 3)?;
103 let k_rot = candle_core::Tensor::cat(&[&(k_half1.broadcast_mul(&cos)? - k_half2.broadcast_mul(&sin)?)?, &(k_half1.broadcast_mul(&sin)? + k_half2.broadcast_mul(&cos)?)?], 3)?;
104 Ok((q_rot, k_rot))
105}
106
107impl Model for GrokModelV2 {
108 type Config = GrokConfig;
109
110 fn new(config: GrokConfig) -> Result<Self> {
111 let device = Device::CPU;
112 let embed_tokens = ops_fn::zeros(&[config.vocab_size, config.hidden_size], DataType::Float32, &device)?;
113 let norm = ops_fn::zeros(&[config.hidden_size], DataType::Float32, &device)?;
114 let lm_head = ops_fn::zeros(&[config.vocab_size, config.hidden_size], DataType::Float32, &device)?;
115 let mut layers = Vec::with_capacity(config.num_hidden_layers);
116 for _ in 0..config.num_hidden_layers { layers.push(GrokLayer::new(&config, &device)?); }
117 Ok(Self { config, device, embed_tokens, layers, norm, lm_head })
118 }
119
120 fn from_weights(config: GrokConfig, weights: ModelWeights) -> Result<Self> {
121 let mut model = Self::new(config)?;
122 if let Some(w) = weights.get("model.embed_tokens.weight") { model.embed_tokens = w.clone(); }
123 if let Some(w) = weights.get("model.norm.weight") { model.norm = w.clone(); }
124 if let Some(w) = weights.get("lm_head.weight") { model.lm_head = w.clone(); }
125 for (i, layer) in model.layers.iter_mut().enumerate() { layer.load_weights(&weights, i)?; }
126 Ok(model)
127 }
128
129 fn forward(&self, inputs: &ModelInputs) -> Result<ModelOutputs> {
130 match inputs {
131 ModelInputs::Text { input_ids, .. } => {
132 let seq_len = input_ids.shape()[1];
133 let mut hidden = ops_fn::embedding(input_ids, &self.embed_tokens)?;
134 for layer in &self.layers { hidden = layer.forward(&hidden, seq_len, self.config.rope_theta)?; }
135 hidden = ops_fn::rms_norm(&hidden, &self.norm, self.config.rms_norm_eps)?;
136 let logits = ops_fn::matmul(&hidden, &ops_fn::transpose(&self.lm_head)?)?;
137 Ok(ModelOutputs::Logits { logits, hidden_states: None })
138 }
139 _ => Err(anyhow::anyhow!("Grok only supports text inputs")),
140 }
141 }
142
143 fn generate(&self, prompt: &str, config: &GenerationConfig) -> Result<String> {
144 use crate::tokenizer::Tokenizer;
145 use rand::Rng;
146 let tokenizer = Tokenizer::new();
147 let mut tokens: Vec<u32> = tokenizer.encode(prompt);
148 for _ in 0..config.max_new_tokens {
149 let tokens_i64: Vec<i64> = tokens.iter().map(|&t| t as i64).collect();
150 let input = Tensor::from_i64_slice(&tokens_i64, &[1, tokens.len()], &self.device)?;
151 let outputs = self.forward(&ModelInputs::text(input))?;
152 let logits = match outputs { ModelOutputs::Logits { logits, .. } => logits, _ => return Err(anyhow::anyhow!("Expected logits")) };
153 let logits_candle = logits.to_candle()?;
154 let last = logits_candle.narrow(1, logits_candle.dims()[1] - 1, 1)?.squeeze(1)?.squeeze(0)?;
155 let logits_vec: Vec<f32> = last.to_vec1()?;
156 let next = if config.do_sample && config.temperature > 0.0 {
157 let scaled: Vec<f32> = logits_vec.iter().map(|&x| x / config.temperature).collect();
158 let max_v = scaled.iter().cloned().fold(f32::NEG_INFINITY, f32::max);
159 let exp_sum: f32 = scaled.iter().map(|&x| (x - max_v).exp()).sum();
160 let probs: Vec<f32> = scaled.iter().map(|&x| (x - max_v).exp() / exp_sum).collect();
161 let mut rng = rand::thread_rng();
162 let r: f32 = rng.gen();
163 let mut cum = 0.0;
164 let mut s = 0u32;
165 for (i, &p) in probs.iter().enumerate() { cum += p; if r <= cum { s = i as u32; break; } }
166 s
167 } else {
168 logits_vec.iter().enumerate().max_by(|a, b| a.1.partial_cmp(b.1).unwrap()).map(|(i, _)| i as u32).unwrap_or(0)
169 };
170 if next == config.eos_token_id { break; }
171 tokens.push(next);
172 }
173 Ok(tokenizer.decode(&tokens))
174 }
175
176 fn config(&self) -> &Self::Config { &self.config }
177 fn memory_requirements(&self) -> MemoryRequirements {
178 let p = self.config.vocab_size * self.config.hidden_size + self.config.num_hidden_layers * 8 * self.config.hidden_size.pow(2);
179 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 }
180 }
181 fn to_device(&mut self, device: &Device) -> Result<()> {
182 self.embed_tokens = self.embed_tokens.to_device(device)?;
183 self.norm = self.norm.to_device(device)?;
184 self.lm_head = self.lm_head.to_device(device)?;
185 for layer in &mut self.layers { layer.to_device(device)?; }
186 self.device = device.clone();
187 Ok(())
188 }
189}
190
191impl GrokLayer {
192 fn new(config: &GrokConfig, device: &Device) -> Result<Self> {
193 Ok(Self {
194 self_attn: GrokAttention::new(config, device)?,
195 moe: GrokMoE::new(config, device)?,
196 input_layernorm: ops_fn::zeros(&[config.hidden_size], DataType::Float32, device)?,
197 post_attention_layernorm: ops_fn::zeros(&[config.hidden_size], DataType::Float32, device)?,
198 })
199 }
200
201 fn forward(&self, hidden_states: &Tensor, seq_len: usize, rope_theta: f32) -> Result<Tensor> {
202 let residual = hidden_states.clone();
203 let h = ops_fn::rms_norm(hidden_states, &self.input_layernorm, 1e-5)?;
204 let attn_out = self.self_attn.forward(&h, seq_len, rope_theta)?;
205 let h = ops_fn::add(&residual, &attn_out)?;
206 let residual = h.clone();
207 let h = ops_fn::rms_norm(&h, &self.post_attention_layernorm, 1e-5)?;
208 let moe_out = self.moe.forward(&h)?;
209 ops_fn::add(&residual, &moe_out)
210 }
211
212 fn load_weights(&mut self, weights: &ModelWeights, idx: usize) -> Result<()> {
213 let p = format!("model.layers.{}", idx);
214 if let Some(w) = weights.get(&format!("{}.self_attn.q_proj.weight", p)) { self.self_attn.q_proj = ops_fn::transpose(w)?; }
215 if let Some(w) = weights.get(&format!("{}.self_attn.k_proj.weight", p)) { self.self_attn.k_proj = ops_fn::transpose(w)?; }
216 if let Some(w) = weights.get(&format!("{}.self_attn.v_proj.weight", p)) { self.self_attn.v_proj = ops_fn::transpose(w)?; }
217 if let Some(w) = weights.get(&format!("{}.self_attn.o_proj.weight", p)) { self.self_attn.o_proj = ops_fn::transpose(w)?; }
218 if let Some(w) = weights.get(&format!("{}.input_layernorm.weight", p)) { self.input_layernorm = w.clone(); }
219 if let Some(w) = weights.get(&format!("{}.post_attention_layernorm.weight", p)) { self.post_attention_layernorm = w.clone(); }
220 Ok(())
221 }
222
223 fn to_device(&mut self, device: &Device) -> Result<()> {
224 self.self_attn.to_device(device)?;
225 self.moe.to_device(device)?;
226 self.input_layernorm = self.input_layernorm.to_device(device)?;
227 self.post_attention_layernorm = self.post_attention_layernorm.to_device(device)?;
228 Ok(())
229 }
230}
231
232impl GrokAttention {
233 fn new(config: &GrokConfig, device: &Device) -> Result<Self> {
234 let head_dim = config.hidden_size / config.num_attention_heads;
235 Ok(Self {
236 q_proj: ops_fn::zeros(&[config.hidden_size, config.num_attention_heads * head_dim], DataType::Float32, device)?,
237 k_proj: ops_fn::zeros(&[config.hidden_size, config.num_key_value_heads * head_dim], DataType::Float32, device)?,
238 v_proj: ops_fn::zeros(&[config.hidden_size, config.num_key_value_heads * head_dim], DataType::Float32, device)?,
239 o_proj: ops_fn::zeros(&[config.num_attention_heads * head_dim, config.hidden_size], DataType::Float32, device)?,
240 num_heads: config.num_attention_heads, num_kv_heads: config.num_key_value_heads, head_dim,
241 scale: 1.0 / (head_dim as f32).sqrt(),
242 })
243 }
244
245 fn forward(&self, hidden_states: &Tensor, seq_len: usize, rope_theta: f32) -> Result<Tensor> {
246 let shape = hidden_states.shape();
247 let batch = shape[0];
248 let q = ops_fn::matmul(hidden_states, &self.q_proj)?.to_candle()?.reshape(&[batch, seq_len, self.num_heads, self.head_dim])?.transpose(1, 2)?;
249 let k = ops_fn::matmul(hidden_states, &self.k_proj)?.to_candle()?.reshape(&[batch, seq_len, self.num_kv_heads, self.head_dim])?.transpose(1, 2)?;
250 let v = ops_fn::matmul(hidden_states, &self.v_proj)?.to_candle()?.reshape(&[batch, seq_len, self.num_kv_heads, self.head_dim])?.transpose(1, 2)?;
251 let (q, k) = apply_rope_grok(&q, &k, seq_len, self.head_dim, rope_theta)?;
252 let num_groups = self.num_heads / self.num_kv_heads;
253 let (k, v) = if num_groups > 1 {
254 (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])?,
255 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])?)
256 } else { (k, v) };
257 let q = q.contiguous()?;
258 let k_t = k.transpose(2, 3)?.contiguous()?;
259 let scores = (q.matmul(&k_t)? * (self.scale as f64))?;
260 let device = scores.device();
261 let mut m = vec![0.0f32; seq_len * seq_len];
262 for i in 0..seq_len { for j in (i+1)..seq_len { m[i*seq_len+j] = f32::NEG_INFINITY; } }
263 let mask = candle_core::Tensor::from_vec(m, &[1, 1, seq_len, seq_len], device)?;
264 let scores = scores.broadcast_add(&mask)?;
265 let v = v.contiguous()?;
266 let attn = candle_nn::ops::softmax_last_dim(&scores)?.matmul(&v)?;
267 let out = attn.transpose(1, 2)?.reshape(&[batch, seq_len, self.num_heads * self.head_dim])?;
268 ops_fn::matmul(&Tensor::from_candle(out), &self.o_proj)
269 }
270
271 fn to_device(&mut self, device: &Device) -> Result<()> {
272 self.q_proj = self.q_proj.to_device(device)?;
273 self.k_proj = self.k_proj.to_device(device)?;
274 self.v_proj = self.v_proj.to_device(device)?;
275 self.o_proj = self.o_proj.to_device(device)?;
276 Ok(())
277 }
278}
279
280impl GrokMoE {
281 fn new(config: &GrokConfig, device: &Device) -> Result<Self> {
282 let router = ops_fn::zeros(&[config.hidden_size, config.num_experts], DataType::Float32, device)?;
283 let mut experts = Vec::with_capacity(config.num_experts);
284 for _ in 0..config.num_experts { experts.push(GrokExpert::new(config, device)?); }
285 Ok(Self { router, experts, num_experts_per_tok: config.num_experts_per_tok })
286 }
287
288 fn forward(&self, hidden_states: &Tensor) -> Result<Tensor> {
289 let shape = hidden_states.shape();
290 let (batch_size, seq_len, hidden_size) = (shape[0], shape[1], shape[2]);
291 let num_tokens = batch_size * seq_len;
292 let k = self.num_experts_per_tok;
293 let flat_hidden = hidden_states.reshape(&[num_tokens, hidden_size])?;
294 let router_logits = ops_fn::matmul(&flat_hidden, &self.router)?;
295 let (topk_weights, topk_indices) = ops_fn::topk(&router_logits, k, -1)?;
296 let routing_weights = ops_fn::softmax(&topk_weights, -1)?;
297 let all_indices: Vec<i64> = topk_indices.to_candle()?.flatten_all()?.to_vec1()?;
298 let all_weights: Vec<f32> = routing_weights.to_candle()?.flatten_all()?.to_vec1()?;
299 let flat_hidden_candle = flat_hidden.to_candle()?;
300 let mut output_data = vec![0.0f32; num_tokens * hidden_size];
301 for tok_idx in 0..num_tokens {
302 let token_hidden = flat_hidden_candle.get(tok_idx)?;
303 let token_tensor = Tensor::from_candle(token_hidden.unsqueeze(0)?);
304 let start = tok_idx * k;
305 let indices = &all_indices[start..start + k];
306 let weights = &all_weights[start..start + k];
307 let mut token_output = ops_fn::zeros(&[1, hidden_size], hidden_states.dtype(), hidden_states.device())?;
308 for (i, &expert_idx) in indices.iter().enumerate() {
309 let expert = &self.experts[expert_idx as usize];
310 let expert_output = expert.forward(&token_tensor)?;
311 let scaled_output = ops_fn::scale(&expert_output, weights[i])?;
312 token_output = ops_fn::add(&token_output, &scaled_output)?;
313 }
314 let token_data: Vec<f32> = token_output.to_candle()?.flatten_all()?.to_vec1()?;
315 for (i, &v) in token_data.iter().enumerate() { output_data[tok_idx * hidden_size + i] = v; }
316 }
317 let output = Tensor::from_f32_slice(&output_data, &[num_tokens, hidden_size], hidden_states.device())?;
318 output.reshape(&[batch_size, seq_len, hidden_size])
319 }
320
321 fn to_device(&mut self, device: &Device) -> Result<()> {
322 self.router = self.router.to_device(device)?;
323 for expert in &mut self.experts { expert.to_device(device)?; }
324 Ok(())
325 }
326}
327
328impl GrokExpert {
329 fn new(config: &GrokConfig, device: &Device) -> Result<Self> {
330 Ok(Self {
331 gate_proj: ops_fn::zeros(&[config.hidden_size, config.intermediate_size], DataType::Float32, device)?,
332 up_proj: ops_fn::zeros(&[config.hidden_size, config.intermediate_size], DataType::Float32, device)?,
333 down_proj: ops_fn::zeros(&[config.intermediate_size, config.hidden_size], DataType::Float32, device)?,
334 })
335 }
336 fn forward(&self, x: &Tensor) -> Result<Tensor> {
337 let gate = ops_fn::matmul(x, &self.gate_proj)?;
338 let up = ops_fn::matmul(x, &self.up_proj)?;
339 let h = ops_fn::mul(&ops_fn::gelu(&gate)?, &up)?;
340 ops_fn::matmul(&h, &self.down_proj)
341 }
342 fn to_device(&mut self, device: &Device) -> Result<()> {
343 self.gate_proj = self.gate_proj.to_device(device)?;
344 self.up_proj = self.up_proj.to_device(device)?;
345 self.down_proj = self.down_proj.to_device(device)?;
346 Ok(())
347 }
348}
349
350#[cfg(test)]
351mod tests {
352 use super::*;
353 #[test]
354 fn test_grok_creation() {
355 let config = GrokConfig { vocab_size: 1000, hidden_size: 64, intermediate_size: 256, num_hidden_layers: 2, num_attention_heads: 4, num_key_value_heads: 2, num_experts: 4, num_experts_per_tok: 2, ..Default::default() };
356 let model = GrokModelV2::new(config).unwrap();
357 assert_eq!(model.config().vocab_size(), 1000);
358 }
359 #[test]
360 fn test_grok_forward() {
361 let config = GrokConfig { vocab_size: 100, hidden_size: 64, intermediate_size: 256, num_hidden_layers: 1, num_attention_heads: 4, num_key_value_heads: 2, num_experts: 4, num_experts_per_tok: 2, ..Default::default() };
362 let model = GrokModelV2::new(config).unwrap();
363 let inputs = ModelInputs::text(ops_fn::zeros(&[1, 4], DataType::Int64, &Device::CPU).unwrap());
364 match model.forward(&inputs).unwrap() { ModelOutputs::Logits { logits, .. } => assert_eq!(logits.shape(), &[1, 4, 100]), _ => panic!() }
365 }
366}