use crate::model_config;
use super::traits::*;
use anyhow::Result;
use serde::{Serialize, Deserialize};
model_config!(GraniteConfig {
vocab_size: usize = 32000,
hidden_size: usize = 4096,
intermediate_size: usize = 11008,
num_hidden_layers: usize = 32,
num_attention_heads: usize = 32,
num_key_value_heads: usize = 8,
hidden_act: String = "silu".to_string(),
max_position_embeddings: usize = 8192,
initializer_range: f32 = 0.02,
rms_norm_eps: f32 = 1e-5,
use_cache: bool = true,
pad_token_id: i64 = 0,
bos_token_id: i64 = 1,
eos_token_id: i64 = 2,
tie_word_embeddings: bool = false,
rope_theta: f32 = 10000.0,
attention_multiplier: f32 = 1.0,
residual_multiplier: f32 = 1.0,
});
impl GraniteConfig {
pub fn from_gguf_config(gguf: &crate::weight_loader_core::GGUFModelConfig) -> Self {
Self {
vocab_size: gguf.vocab_size,
hidden_size: gguf.hidden_size,
intermediate_size: gguf.intermediate_size,
num_hidden_layers: gguf.num_hidden_layers,
num_attention_heads: gguf.num_attention_heads,
num_key_value_heads: gguf.num_key_value_heads,
max_position_embeddings: gguf.max_position_embeddings,
rope_theta: gguf.rope_theta,
..Default::default()
}
}
}
pub struct GraniteModelV2 {
config: GraniteConfig,
device: Device,
embed_tokens: Tensor,
layers: Vec<GraniteLayer>,
norm: Tensor,
lm_head: Tensor,
}
pub struct GraniteLayer {
self_attn: GraniteAttention,
mlp: GraniteMLP,
input_layernorm: Tensor,
post_attention_layernorm: Tensor,
}
pub struct GraniteAttention {
q_proj: Tensor,
k_proj: Tensor,
v_proj: Tensor,
o_proj: Tensor,
num_heads: usize,
num_kv_heads: usize,
head_dim: usize,
scale: f32,
}
pub struct GraniteMLP {
gate_proj: Tensor,
up_proj: Tensor,
down_proj: Tensor,
}
fn apply_rope_granite(
q: &candle_core::Tensor,
k: &candle_core::Tensor,
seq_len: usize,
head_dim: usize,
rope_theta: f32,
) -> Result<(candle_core::Tensor, candle_core::Tensor)> {
let device = q.device();
let half_dim = head_dim / 2;
let inv_freq: Vec<f32> = (0..half_dim)
.map(|i| 1.0 / rope_theta.powf((2 * i) as f32 / head_dim as f32))
.collect();
let positions: Vec<f32> = (0..seq_len).map(|p| p as f32).collect();
let mut angles = Vec::with_capacity(seq_len * half_dim);
for pos in &positions {
for freq in &inv_freq {
angles.push(pos * freq);
}
}
let angles_tensor = candle_core::Tensor::from_vec(angles, &[seq_len, half_dim], device)?;
let cos = angles_tensor.cos()?.unsqueeze(0)?.unsqueeze(0)?;
let sin = angles_tensor.sin()?.unsqueeze(0)?.unsqueeze(0)?;
let q_half1 = q.narrow(3, 0, half_dim)?;
let q_half2 = q.narrow(3, half_dim, half_dim)?;
let k_half1 = k.narrow(3, 0, half_dim)?;
let k_half2 = k.narrow(3, half_dim, half_dim)?;
let q_rot1 = (q_half1.broadcast_mul(&cos)? - q_half2.broadcast_mul(&sin)?)?;
let q_rot2 = (q_half1.broadcast_mul(&sin)? + q_half2.broadcast_mul(&cos)?)?;
let k_rot1 = (k_half1.broadcast_mul(&cos)? - k_half2.broadcast_mul(&sin)?)?;
let k_rot2 = (k_half1.broadcast_mul(&sin)? + k_half2.broadcast_mul(&cos)?)?;
Ok((
candle_core::Tensor::cat(&[&q_rot1, &q_rot2], 3)?,
candle_core::Tensor::cat(&[&k_rot1, &k_rot2], 3)?
))
}
impl Model for GraniteModelV2 {
type Config = GraniteConfig;
fn new(config: GraniteConfig) -> Result<Self> {
let device = Device::CPU;
let embed_tokens = ops_fn::zeros(&[config.vocab_size, config.hidden_size], DataType::Float32, &device)?;
let norm = ops_fn::zeros(&[config.hidden_size], DataType::Float32, &device)?;
let lm_head = ops_fn::zeros(&[config.vocab_size, config.hidden_size], DataType::Float32, &device)?;
let mut layers = Vec::with_capacity(config.num_hidden_layers);
for _ in 0..config.num_hidden_layers {
layers.push(GraniteLayer::new(&config, &device)?);
}
Ok(Self { config, device, embed_tokens, layers, norm, lm_head })
}
fn from_weights(config: GraniteConfig, weights: ModelWeights) -> Result<Self> {
let mut model = Self::new(config)?;
if let Some(w) = weights.get("model.embed_tokens.weight") { model.embed_tokens = w.clone(); }
if let Some(w) = weights.get("model.norm.weight") { model.norm = w.clone(); }
if model.config.tie_word_embeddings {
model.lm_head = model.embed_tokens.clone();
} else if let Some(w) = weights.get("lm_head.weight") {
model.lm_head = w.clone();
}
for (i, layer) in model.layers.iter_mut().enumerate() { layer.load_weights(&weights, i)?; }
Ok(model)
}
fn forward(&self, inputs: &ModelInputs) -> Result<ModelOutputs> {
match inputs {
ModelInputs::Text { input_ids, .. } => {
let seq_len = input_ids.shape()[1];
let mut hidden = ops_fn::embedding(input_ids, &self.embed_tokens)?;
for layer in &self.layers {
hidden = layer.forward(&hidden, seq_len, self.config.rope_theta, self.config.residual_multiplier)?;
}
hidden = ops_fn::rms_norm(&hidden, &self.norm, self.config.rms_norm_eps)?;
let logits = ops_fn::matmul(&hidden, &ops_fn::transpose(&self.lm_head)?)?;
Ok(ModelOutputs::Logits { logits, hidden_states: None })
}
_ => Err(anyhow::anyhow!("Granite only supports text inputs")),
}
}
fn generate(&self, prompt: &str, config: &GenerationConfig) -> Result<String> {
use crate::tokenizer::Tokenizer;
use rand::Rng;
let tokenizer = Tokenizer::new();
let mut tokens: Vec<u32> = tokenizer.encode(prompt);
for _ in 0..config.max_new_tokens {
let tokens_i64: Vec<i64> = tokens.iter().map(|&t| t as i64).collect();
let input = Tensor::from_i64_slice(&tokens_i64, &[1, tokens.len()], &self.device)?;
let outputs = self.forward(&ModelInputs::text(input))?;
let logits = match outputs { ModelOutputs::Logits { logits, .. } => logits, _ => return Err(anyhow::anyhow!("Expected logits")) };
let logits_candle = logits.to_candle()?;
let last = logits_candle.narrow(1, logits_candle.dims()[1] - 1, 1)?.squeeze(1)?.squeeze(0)?;
let logits_vec: Vec<f32> = last.to_vec1()?;
let next = if config.do_sample && config.temperature > 0.0 {
let scaled: Vec<f32> = logits_vec.iter().map(|&x| x / config.temperature).collect();
let max_v = scaled.iter().cloned().fold(f32::NEG_INFINITY, f32::max);
let exp_sum: f32 = scaled.iter().map(|&x| (x - max_v).exp()).sum();
let probs: Vec<f32> = scaled.iter().map(|&x| (x - max_v).exp() / exp_sum).collect();
let mut rng = rand::thread_rng();
let r: f32 = rng.gen();
let mut cum = 0.0;
let mut s = 0u32;
for (i, &p) in probs.iter().enumerate() { cum += p; if r <= cum { s = i as u32; break; } }
s
} else {
logits_vec.iter().enumerate().max_by(|a, b| a.1.partial_cmp(b.1).unwrap()).map(|(i, _)| i as u32).unwrap_or(0)
};
if next == config.eos_token_id { break; }
tokens.push(next);
}
Ok(tokenizer.decode(&tokens))
}
fn config(&self) -> &Self::Config { &self.config }
fn memory_requirements(&self) -> MemoryRequirements {
let p = self.config.vocab_size * self.config.hidden_size + self.config.num_hidden_layers * 8 * self.config.hidden_size.pow(2);
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 }
}
fn to_device(&mut self, device: &Device) -> Result<()> {
self.embed_tokens = self.embed_tokens.to_device(device)?;
self.norm = self.norm.to_device(device)?;
self.lm_head = self.lm_head.to_device(device)?;
for layer in &mut self.layers { layer.to_device(device)?; }
self.device = device.clone();
Ok(())
}
}
impl GraniteLayer {
fn new(config: &GraniteConfig, device: &Device) -> Result<Self> {
Ok(Self {
self_attn: GraniteAttention::new(config, device)?,
mlp: GraniteMLP::new(config, device)?,
input_layernorm: ops_fn::zeros(&[config.hidden_size], DataType::Float32, device)?,
post_attention_layernorm: ops_fn::zeros(&[config.hidden_size], DataType::Float32, device)?,
})
}
fn forward(&self, hidden_states: &Tensor, seq_len: usize, rope_theta: f32, residual_multiplier: f32) -> Result<Tensor> {
let residual = hidden_states.clone();
let h = ops_fn::rms_norm(hidden_states, &self.input_layernorm, 1e-5)?;
let attn_out = self.self_attn.forward(&h, seq_len, rope_theta)?;
let attn_scaled = if residual_multiplier != 1.0 {
ops_fn::scale(&attn_out, residual_multiplier)?
} else {
attn_out
};
let h = ops_fn::add(&residual, &attn_scaled)?;
let residual = h.clone();
let h = ops_fn::rms_norm(&h, &self.post_attention_layernorm, 1e-5)?;
let mlp_out = self.mlp.forward(&h)?;
let mlp_scaled = if residual_multiplier != 1.0 {
ops_fn::scale(&mlp_out, residual_multiplier)?
} else {
mlp_out
};
ops_fn::add(&residual, &mlp_scaled)
}
fn load_weights(&mut self, weights: &ModelWeights, idx: usize) -> Result<()> {
let p = format!("model.layers.{}", idx);
if let Some(w) = weights.get(&format!("{}.self_attn.q_proj.weight", p)) { self.self_attn.q_proj = ops_fn::transpose(w)?; }
if let Some(w) = weights.get(&format!("{}.self_attn.k_proj.weight", p)) { self.self_attn.k_proj = ops_fn::transpose(w)?; }
if let Some(w) = weights.get(&format!("{}.self_attn.v_proj.weight", p)) { self.self_attn.v_proj = ops_fn::transpose(w)?; }
if let Some(w) = weights.get(&format!("{}.self_attn.o_proj.weight", p)) { self.self_attn.o_proj = ops_fn::transpose(w)?; }
if let Some(w) = weights.get(&format!("{}.mlp.gate_proj.weight", p)) { self.mlp.gate_proj = ops_fn::transpose(w)?; }
if let Some(w) = weights.get(&format!("{}.mlp.up_proj.weight", p)) { self.mlp.up_proj = ops_fn::transpose(w)?; }
if let Some(w) = weights.get(&format!("{}.mlp.down_proj.weight", p)) { self.mlp.down_proj = ops_fn::transpose(w)?; }
if let Some(w) = weights.get(&format!("{}.input_layernorm.weight", p)) { self.input_layernorm = w.clone(); }
if let Some(w) = weights.get(&format!("{}.post_attention_layernorm.weight", p)) { self.post_attention_layernorm = w.clone(); }
Ok(())
}
fn to_device(&mut self, device: &Device) -> Result<()> {
self.self_attn.to_device(device)?;
self.mlp.to_device(device)?;
self.input_layernorm = self.input_layernorm.to_device(device)?;
self.post_attention_layernorm = self.post_attention_layernorm.to_device(device)?;
Ok(())
}
}
impl GraniteAttention {
fn new(config: &GraniteConfig, device: &Device) -> Result<Self> {
let head_dim = config.hidden_size / config.num_attention_heads;
Ok(Self {
q_proj: ops_fn::zeros(&[config.hidden_size, config.num_attention_heads * head_dim], DataType::Float32, device)?,
k_proj: ops_fn::zeros(&[config.hidden_size, config.num_key_value_heads * head_dim], DataType::Float32, device)?,
v_proj: ops_fn::zeros(&[config.hidden_size, config.num_key_value_heads * head_dim], DataType::Float32, device)?,
o_proj: ops_fn::zeros(&[config.num_attention_heads * head_dim, config.hidden_size], DataType::Float32, device)?,
num_heads: config.num_attention_heads,
num_kv_heads: config.num_key_value_heads,
head_dim,
scale: 1.0 / (head_dim as f32).sqrt(),
})
}
fn forward(&self, hidden_states: &Tensor, seq_len: usize, rope_theta: f32) -> Result<Tensor> {
let shape = hidden_states.shape();
let batch = shape[0];
let q = ops_fn::matmul(hidden_states, &self.q_proj)?.to_candle()?;
let k = ops_fn::matmul(hidden_states, &self.k_proj)?.to_candle()?;
let v = ops_fn::matmul(hidden_states, &self.v_proj)?.to_candle()?;
let q = q.reshape(&[batch, seq_len, self.num_heads, self.head_dim])?.transpose(1, 2)?;
let k = k.reshape(&[batch, seq_len, self.num_kv_heads, self.head_dim])?.transpose(1, 2)?;
let v = v.reshape(&[batch, seq_len, self.num_kv_heads, self.head_dim])?.transpose(1, 2)?;
let (q, k) = apply_rope_granite(&q, &k, seq_len, self.head_dim, rope_theta)?;
let num_groups = self.num_heads / self.num_kv_heads;
let (k, v) = if num_groups > 1 {
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])?;
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])?;
(k_exp, v_exp)
} else {
(k, v)
};
let q = q.contiguous()?;
let k_t = k.transpose(2, 3)?.contiguous()?;
let scores = (q.matmul(&k_t)? * (self.scale as f64))?;
let device = scores.device();
let mask = {
let mut m = vec![0.0f32; seq_len * seq_len];
for i in 0..seq_len { for j in (i+1)..seq_len { m[i*seq_len+j] = f32::NEG_INFINITY; } }
candle_core::Tensor::from_vec(m, &[1, 1, seq_len, seq_len], device)?
};
let scores = scores.broadcast_add(&mask)?;
let v = v.contiguous()?;
let attn = candle_nn::ops::softmax_last_dim(&scores)?.matmul(&v)?;
let out = attn.transpose(1, 2)?.reshape(&[batch, seq_len, self.num_heads * self.head_dim])?;
ops_fn::matmul(&Tensor::from_candle(out), &self.o_proj)
}
fn to_device(&mut self, device: &Device) -> Result<()> {
self.q_proj = self.q_proj.to_device(device)?;
self.k_proj = self.k_proj.to_device(device)?;
self.v_proj = self.v_proj.to_device(device)?;
self.o_proj = self.o_proj.to_device(device)?;
Ok(())
}
}
impl GraniteMLP {
fn new(config: &GraniteConfig, device: &Device) -> Result<Self> {
Ok(Self {
gate_proj: ops_fn::zeros(&[config.hidden_size, config.intermediate_size], DataType::Float32, device)?,
up_proj: ops_fn::zeros(&[config.hidden_size, config.intermediate_size], DataType::Float32, device)?,
down_proj: ops_fn::zeros(&[config.intermediate_size, config.hidden_size], DataType::Float32, device)?,
})
}
fn forward(&self, x: &Tensor) -> Result<Tensor> {
let gate = ops_fn::matmul(x, &self.gate_proj)?;
let up = ops_fn::matmul(x, &self.up_proj)?;
let h = ops_fn::mul(&ops_fn::silu(&gate)?, &up)?;
ops_fn::matmul(&h, &self.down_proj)
}
fn to_device(&mut self, device: &Device) -> Result<()> {
self.gate_proj = self.gate_proj.to_device(device)?;
self.up_proj = self.up_proj.to_device(device)?;
self.down_proj = self.down_proj.to_device(device)?;
Ok(())
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_granite_creation() {
let config = GraniteConfig { vocab_size: 1000, hidden_size: 128, intermediate_size: 512, num_hidden_layers: 2, num_attention_heads: 4, num_key_value_heads: 4, ..Default::default() };
let model = GraniteModelV2::new(config).unwrap();
assert_eq!(model.config().vocab_size(), 1000);
}
#[test]
fn test_granite_forward() {
let config = GraniteConfig { vocab_size: 100, hidden_size: 64, intermediate_size: 256, num_hidden_layers: 1, num_attention_heads: 4, num_key_value_heads: 4, ..Default::default() };
let model = GraniteModelV2::new(config).unwrap();
let inputs = ModelInputs::text(ops_fn::zeros(&[2, 8], DataType::Int64, &Device::CPU).unwrap());
match model.forward(&inputs).unwrap() { ModelOutputs::Logits { logits, .. } => assert_eq!(logits.shape(), &[2, 8, 100]), _ => panic!() }
}
}