embed_anything 0.7.1

//! Gemma LLM architecture (Google) inference implementation.
//!
//! See ["Introducing Gemma 3: The most capable model you can run on a single GPU or TPU"](https://blog.google/technology/developers/gemma-3/)
//!
//! Based on implementations from HuggingFace transformers.

use std::sync::Arc;

use candle_core::{DType, Device, Module, Result, Tensor, D};
use candle_nn::{linear_b as linear, Activation, Linear, VarBuilder};

use crate::models::qwen3::repeat_kv;

fn default_rope_local_base_freq() -> f64 {
    10000.0
}

#[derive(serde::Deserialize, Debug, Clone)]
pub struct Config {
    pub attention_bias: bool,
    pub head_dim: usize,
    pub hidden_activation: Activation,
    pub hidden_size: usize,
    pub intermediate_size: usize,
    pub num_attention_heads: usize,
    pub num_hidden_layers: usize,
    pub num_key_value_heads: usize,
    pub rms_norm_eps: f64,
    pub rope_theta: f64,
    #[serde(default = "default_rope_local_base_freq")]
    pub rope_local_base_freq: f64,
    pub vocab_size: usize,
    pub final_logit_softcapping: Option<f64>,
    pub attn_logit_softcapping: Option<f64>,
    pub query_pre_attn_scalar: usize,
    pub sliding_window: usize,
    pub _sliding_window_pattern: usize,
    pub max_position_embeddings: usize,
}

#[derive(Debug, Clone)]
struct RmsNorm {
    weight: Tensor,
    eps: f64,
}

impl RmsNorm {
    fn new(dim: usize, eps: f64, vb: VarBuilder) -> Result<Self> {
        let weight = vb.get(dim, "weight")?;
        Ok(Self { weight, eps })
    }
}

impl Module for RmsNorm {
    fn forward(&self, x: &Tensor) -> Result<Tensor> {
        let x_dtype = x.dtype();
        let internal_dtype = match x_dtype {
            DType::F16 | DType::BF16 => DType::F32,
            d => d,
        };
        let hidden_size = x.dim(D::Minus1)?;
        let x = x.to_dtype(internal_dtype)?;
        let norm_x = (x.sqr()?.sum_keepdim(D::Minus1)? / hidden_size as f64)?;
        let x_normed = x.broadcast_div(&(norm_x + self.eps)?.sqrt()?)?;
        x_normed
            .to_dtype(x_dtype)?
            .broadcast_mul(&(&self.weight + 1.0)?)
    }
}

#[derive(Debug, Clone)]
struct RotaryEmbedding {
    sin: Tensor,
    cos: Tensor,
}

impl RotaryEmbedding {
    fn new(dtype: DType, rope_theta: f64, cfg: &Config, dev: &Device) -> Result<Self> {
        let dim = cfg.head_dim;
        let max_seq_len = cfg.max_position_embeddings;
        let inv_freq: Vec<_> = (0..dim)
            .step_by(2)
            .map(|i| 1f32 / rope_theta.powf(i as f64 / dim as f64) as f32)
            .collect();
        let inv_freq_len = inv_freq.len();
        let inv_freq = Tensor::from_vec(inv_freq, (1, inv_freq_len), dev)?.to_dtype(dtype)?;
        let t = Tensor::arange(0u32, max_seq_len as u32, dev)?
            .to_dtype(dtype)?
            .reshape((max_seq_len, 1))?;
        let freqs = t.matmul(&inv_freq)?;
        Ok(Self {
            sin: freqs.sin()?,
            cos: freqs.cos()?,
        })
    }

    fn apply_rotary_emb_qkv(
        &self,
        q: &Tensor,
        k: &Tensor,
        seqlen_offset: usize,
    ) -> Result<(Tensor, Tensor)> {
        let (_b_sz, _h, seq_len, _n_embd) = q.dims4()?;
        let cos = self.cos.narrow(0, seqlen_offset, seq_len)?;
        let sin = self.sin.narrow(0, seqlen_offset, seq_len)?;
        let q_embed = candle_nn::rotary_emb::rope(&q.contiguous()?, &cos, &sin)?;
        let k_embed = candle_nn::rotary_emb::rope(&k.contiguous()?, &cos, &sin)?;
        Ok((q_embed, k_embed))
    }
}

#[derive(Debug, Clone)]
#[allow(clippy::upper_case_acronyms)]
struct MLP {
    gate_proj: Linear,
    up_proj: Linear,
    down_proj: Linear,
    act_fn: candle_nn::Activation,
}

impl MLP {
    fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
        let hidden_sz = cfg.hidden_size;
        let intermediate_sz = cfg.intermediate_size;
        let gate_proj = linear(hidden_sz, intermediate_sz, false, vb.pp("gate_proj"))?;
        let up_proj = linear(hidden_sz, intermediate_sz, false, vb.pp("up_proj"))?;
        let down_proj = linear(intermediate_sz, hidden_sz, false, vb.pp("down_proj"))?;
        Ok(Self {
            gate_proj,
            up_proj,
            down_proj,
            act_fn: cfg.hidden_activation,
        })
    }
}

impl Module for MLP {
    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
        let lhs = xs.apply(&self.gate_proj)?.apply(&self.act_fn)?;
        let rhs = xs.apply(&self.up_proj)?;
        (lhs * rhs)?.apply(&self.down_proj)
    }
}

#[derive(Debug, Clone)]
enum KvCache {
    Normal(candle_nn::kv_cache::KvCache),
    Rotating(candle_nn::kv_cache::RotatingKvCache),
}

#[derive(Debug, Clone)]
struct Attention {
    q_proj: Linear,
    k_proj: Linear,
    v_proj: Linear,
    o_proj: Linear,
    q_norm: RmsNorm,
    k_norm: RmsNorm,
    num_heads: usize,
    num_kv_heads: usize,
    num_kv_groups: usize,
    head_dim: usize,
    attn_logit_softcapping: Option<f64>,
    rotary_emb: Arc<RotaryEmbedding>,
    kv_cache: KvCache,
    use_flash_attn: bool,
}

impl Attention {
    fn new(
        rotary_emb: Arc<RotaryEmbedding>,
        use_flash_attn: bool,
        is_sliding: bool,
        cfg: &Config,
        vb: VarBuilder,
    ) -> Result<Self> {
        let hidden_sz = cfg.hidden_size;
        let num_heads = cfg.num_attention_heads;
        let num_kv_heads = cfg.num_key_value_heads;
        let num_kv_groups = num_heads / num_kv_heads;
        let head_dim = cfg.head_dim;
        let bias = cfg.attention_bias;
        let q_proj = linear(hidden_sz, num_heads * head_dim, bias, vb.pp("q_proj"))?;
        let k_proj = linear(hidden_sz, num_kv_heads * head_dim, bias, vb.pp("k_proj"))?;
        let v_proj = linear(hidden_sz, num_kv_heads * head_dim, bias, vb.pp("v_proj"))?;
        let o_proj = linear(num_heads * head_dim, hidden_sz, bias, vb.pp("o_proj"))?;
        let q_norm = RmsNorm::new(head_dim, cfg.rms_norm_eps, vb.pp("q_norm"))?;
        let k_norm = RmsNorm::new(head_dim, cfg.rms_norm_eps, vb.pp("k_norm"))?;
        let kv_cache = if is_sliding {
            KvCache::Rotating(candle_nn::kv_cache::RotatingKvCache::new(
                2,
                cfg.sliding_window,
            ))
        } else {
            KvCache::Normal(candle_nn::kv_cache::KvCache::new(2, cfg.sliding_window))
        };
        Ok(Self {
            q_proj,
            k_proj,
            v_proj,
            o_proj,
            q_norm,
            k_norm,
            num_heads,
            num_kv_heads,
            num_kv_groups,
            head_dim,
            attn_logit_softcapping: cfg.attn_logit_softcapping,
            rotary_emb,
            kv_cache,
            use_flash_attn,
        })
    }

    fn forward(
        &mut self,
        xs: &Tensor,
        attention_mask: Option<&Tensor>,
        seqlen_offset: usize,
    ) -> Result<Tensor> {
        let (b_sz, q_len, _) = xs.dims3()?;

        let query_states = self.q_proj.forward(xs)?;
        let key_states = self.k_proj.forward(xs)?;
        let value_states = self.v_proj.forward(xs)?;

        let query_states = query_states
            .reshape((b_sz, q_len, self.num_heads, self.head_dim))?
            .transpose(1, 2)?;
        let key_states = key_states
            .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
            .transpose(1, 2)?;
        let value_states = value_states
            .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
            .transpose(1, 2)?;
        let query_states = self.q_norm.forward(&query_states)?;
        let key_states = self.k_norm.forward(&key_states)?;

        let (query_states, key_states) =
            self.rotary_emb
                .apply_rotary_emb_qkv(&query_states, &key_states, seqlen_offset)?;

        let (key_states, value_states) = match &mut self.kv_cache {
            KvCache::Normal(cache) => cache.append(&key_states, &value_states)?,
            KvCache::Rotating(cache) => cache.append(&key_states, &value_states)?,
        };

        let key_states = repeat_kv(key_states, self.num_kv_groups)?.contiguous()?;
        let value_states = repeat_kv(value_states, self.num_kv_groups)?.contiguous()?;

        let attn_output = if self.use_flash_attn {
            // flash-attn expects (b_sz, seq_len, nheads, head_dim)
            let q = query_states.transpose(1, 2)?;
            let k = key_states.transpose(1, 2)?;
            let v = value_states.transpose(1, 2)?;
            let scale = 1f32 / (self.head_dim as f32).sqrt();
            flash_attn(&q, &k, &v, scale, attention_mask.is_some())?.transpose(1, 2)?
        } else {
            let scale = 1f64 / f64::sqrt(self.head_dim as f64);
            let attn_weights = (query_states.matmul(&key_states.transpose(2, 3)?)? * scale)?;

            let attn_weights = match self.attn_logit_softcapping {
                None => attn_weights,
                Some(sc) => ((attn_weights / sc)?.tanh()? * sc)?,
            };

            let attn_weights = match attention_mask {
                None => attn_weights,
                Some(mask) => attn_weights.broadcast_add(mask)?,
            };
            let attn_weights = candle_nn::ops::softmax_last_dim(&attn_weights)?;
            attn_weights.matmul(&value_states)?
        };
        attn_output
            .transpose(1, 2)?
            .reshape((b_sz, q_len, ()))?
            .apply(&self.o_proj)
    }

    fn clear_kv_cache(&mut self) {
        match &mut self.kv_cache {
            KvCache::Normal(c) => c.reset(),
            KvCache::Rotating(c) => c.reset(),
        }
    }
}

#[cfg(feature = "flash-attn")]
fn flash_attn(
    q: &Tensor,
    k: &Tensor,
    v: &Tensor,
    softmax_scale: f32,
    causal: bool,
) -> Result<Tensor> {
    candle_flash_attn::flash_attn(q, k, v, softmax_scale, causal)
}

#[cfg(not(feature = "flash-attn"))]
fn flash_attn(_: &Tensor, _: &Tensor, _: &Tensor, _: f32, _: bool) -> Result<Tensor> {
    unimplemented!("compile with '--features flash-attn'")
}

#[derive(Debug, Clone)]
struct DecoderLayer {
    self_attn: Attention,
    mlp: MLP,
    input_layernorm: RmsNorm,
    pre_feedforward_layernorm: RmsNorm,
    post_feedforward_layernorm: RmsNorm,
    post_attention_layernorm: RmsNorm,
    is_sliding: bool,
}

impl DecoderLayer {
    fn new(
        rotary_emb: Arc<RotaryEmbedding>,
        use_flash_attn: bool,
        is_sliding: bool,
        cfg: &Config,
        vb: VarBuilder,
    ) -> Result<Self> {
        let self_attn = Attention::new(
            rotary_emb,
            use_flash_attn,
            is_sliding,
            cfg,
            vb.pp("self_attn"),
        )?;
        let mlp = MLP::new(cfg, vb.pp("mlp"))?;
        let input_layernorm =
            RmsNorm::new(cfg.hidden_size, cfg.rms_norm_eps, vb.pp("input_layernorm"))?;
        let pre_feedforward_layernorm = RmsNorm::new(
            cfg.hidden_size,
            cfg.rms_norm_eps,
            vb.pp("pre_feedforward_layernorm"),
        )?;
        let post_feedforward_layernorm = RmsNorm::new(
            cfg.hidden_size,
            cfg.rms_norm_eps,
            vb.pp("post_feedforward_layernorm"),
        )?;
        let post_attention_layernorm = RmsNorm::new(
            cfg.hidden_size,
            cfg.rms_norm_eps,
            vb.pp("post_attention_layernorm"),
        )?;
        Ok(Self {
            self_attn,
            mlp,
            input_layernorm,
            pre_feedforward_layernorm,
            post_feedforward_layernorm,
            post_attention_layernorm,
            is_sliding,
        })
    }

    fn forward(
        &mut self,
        xs: &Tensor,
        attention_mask: Option<&Tensor>,
        seqlen_offset: usize,
    ) -> Result<Tensor> {
        let residual = xs;
        let xs = self.input_layernorm.forward(xs)?;
        let xs = self.self_attn.forward(&xs, attention_mask, seqlen_offset)?;
        let xs = xs.apply(&self.post_attention_layernorm)?;
        let xs = (xs + residual)?;
        let residual = &xs;
        let xs = xs.apply(&self.pre_feedforward_layernorm)?;
        let xs = xs.apply(&self.mlp)?;
        let xs = xs.apply(&self.post_feedforward_layernorm)?;
        residual + xs
    }

    fn clear_kv_cache(&mut self) {
        self.self_attn.clear_kv_cache()
    }
}

#[derive(Debug, Clone)]
pub struct Model {
    embed_tokens: candle_nn::Embedding,
    layers: Vec<DecoderLayer>,
    norm: RmsNorm,
    device: Device,
    dtype: DType,
    hidden_size: usize,
    sliding_window: usize,
}

impl Model {
    pub fn new(use_flash_attn: bool, cfg: &Config, vb: VarBuilder) -> Result<Self> {
        let embed_tokens =
            candle_nn::embedding(cfg.vocab_size, cfg.hidden_size, vb.pp("embed_tokens"))?;
        // Gemma3 uses two RoPE base frequencies: a global one (rope_theta) for
        // full-attention layers and a local one (rope_local_base_freq) for
        // sliding-attention layers.
        let global_rotary_emb = Arc::new(RotaryEmbedding::new(
            vb.dtype(),
            cfg.rope_theta,
            cfg,
            vb.device(),
        )?);
        let local_rotary_emb = Arc::new(RotaryEmbedding::new(
            vb.dtype(),
            cfg.rope_local_base_freq,
            cfg,
            vb.device(),
        )?);
        let mut layers = Vec::with_capacity(cfg.num_hidden_layers);
        let vb_l = vb.pp("layers");
        for layer_idx in 0..cfg.num_hidden_layers {
            let is_sliding = (layer_idx + 1) % cfg._sliding_window_pattern > 0;
            let rotary_emb = if is_sliding {
                local_rotary_emb.clone()
            } else {
                global_rotary_emb.clone()
            };
            let layer = DecoderLayer::new(
                rotary_emb,
                use_flash_attn,
                is_sliding,
                cfg,
                vb_l.pp(layer_idx),
            )?;
            layers.push(layer)
        }
        let norm = RmsNorm::new(cfg.hidden_size, cfg.rms_norm_eps, vb.pp("norm"))?;
        Ok(Self {
            embed_tokens,
            layers,
            norm,
            device: vb.device().clone(),
            dtype: vb.dtype(),
            hidden_size: cfg.hidden_size,
            sliding_window: cfg.sliding_window,
        })
    }

    /// Builds an additive attention bias for EmbeddingGemma's bidirectional
    /// attention (`use_bidirectional_attention=true`): a token may attend to every
    /// non-padding token. Sliding-attention layers are additionally restricted to a
    /// symmetric band `|i - j| < window`; full-attention layers pass `window = None`.
    fn bidirectional_mask(
        &self,
        attention_mask: Option<&Tensor>,
        b_size: usize,
        tgt_len: usize,
        window: Option<usize>,
    ) -> Result<Tensor> {
        let real: Vec<Vec<bool>> = match attention_mask {
            Some(am) => am
                .to_vec2::<u32>()?
                .into_iter()
                .map(|row| row.into_iter().map(|v| v != 0).collect())
                .collect(),
            None => vec![vec![true; tgt_len]; b_size],
        };
        let mut data = Vec::with_capacity(b_size * tgt_len * tgt_len);
        for row in real.iter().take(b_size) {
            for i in 0..tgt_len {
                for (j, &is_real) in row.iter().enumerate() {
                    let out_of_band = match window {
                        Some(w) => ((i as i64) - (j as i64)).unsigned_abs() as usize >= w,
                        None => false,
                    };
                    let masked = !is_real || out_of_band;
                    data.push(if masked { f32::NEG_INFINITY } else { 0.0 });
                }
            }
        }
        Tensor::from_vec(data, (b_size, 1, tgt_len, tgt_len), &self.device)?.to_dtype(self.dtype)
    }

    pub fn forward(
        &mut self,
        input_ids: &Tensor,
        attention_mask: Option<&Tensor>,
        seqlen_offset: usize,
    ) -> Result<Tensor> {
        let (b_size, seq_len) = input_ids.dims2()?;
        let (full_mask, sliding_mask) = if seq_len <= 1 {
            (None, None)
        } else {
            (
                Some(self.bidirectional_mask(attention_mask, b_size, seq_len, None)?),
                Some(self.bidirectional_mask(
                    attention_mask,
                    b_size,
                    seq_len,
                    Some(self.sliding_window),
                )?),
            )
        };
        let xs = self.embed_tokens.forward(input_ids)?;
        let mut xs = (xs * (self.hidden_size as f64).sqrt())?;
        for layer in self.layers.iter_mut() {
            let mask = if layer.is_sliding {
                sliding_mask.as_ref()
            } else {
                full_mask.as_ref()
            };
            xs = layer.forward(&xs, mask, seqlen_offset)?;
        }
        xs.apply(&self.norm)
    }

    pub fn clear_kv_cache(&mut self) {
        for layer in self.layers.iter_mut() {
            layer.clear_kv_cache()
        }
    }
}