Gemma 3 initial setup (text only). (#2802)

* Gemma 3 initial setup (text only).

* Use the rotating kv cache for the sliding window.

candle-transformers/src/models/gemma3.rs

@@ -0,0 +1,483 @@
+//! 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::{DType, Device, Module, Result, Tensor, D};
+use candle_nn::{linear_b as linear, Activation, Linear, VarBuilder};
+
+#[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,
+    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, 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 / cfg.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 = crate::utils::repeat_kv(key_states, self.num_kv_groups)?.contiguous()?;
+        let value_states =
+            crate::utils::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,
+}
+
+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,
+        })
+    }
+
+    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,
+    lm_head: Linear,
+    final_logit_softcapping: Option<f64>,
+    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 vb_m = vb.pp("model");
+        let embed_tokens =
+            candle_nn::embedding(cfg.vocab_size, cfg.hidden_size, vb_m.pp("embed_tokens"))?;
+        let rotary_emb = Arc::new(RotaryEmbedding::new(vb.dtype(), cfg, vb_m.device())?);
+        let mut layers = Vec::with_capacity(cfg.num_hidden_layers);
+        let vb_l = vb_m.pp("layers");
+        for layer_idx in 0..cfg.num_hidden_layers {
+            let is_sliding = (layer_idx + 1) % cfg.sliding_window_pattern > 0;
+            let layer = DecoderLayer::new(
+                rotary_emb.clone(),
+                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_m.pp("norm"))?;
+        let lm_head = Linear::new(embed_tokens.embeddings().clone(), None);
+        Ok(Self {
+            embed_tokens,
+            layers,
+            norm,
+            lm_head,
+            final_logit_softcapping: cfg.final_logit_softcapping,
+            device: vb.device().clone(),
+            dtype: vb.dtype(),
+            hidden_size: cfg.hidden_size,
+            sliding_window: cfg.sliding_window,
+        })
+    }
+
+    fn prepare_decoder_attention_mask(
+        &self,
+        b_size: usize,
+        tgt_len: usize,
+        seqlen_offset: usize,
+    ) -> Result<Tensor> {
+        let mask: Vec<_> = match Some(self.sliding_window) {
+            None => (0..tgt_len)
+                .flat_map(|i| (0..tgt_len).map(move |j| if i < j { f32::NEG_INFINITY } else { 0. }))
+                .collect(),
+            Some(sliding_window) => (0..tgt_len)
+                .flat_map(|i| {
+                    (0..tgt_len).map(move |j| {
+                        if i < j || j + sliding_window < i {
+                            f32::NEG_INFINITY
+                        } else {
+                            0.
+                        }
+                    })
+                })
+                .collect(),
+        };
+        let mask = Tensor::from_slice(&mask, (tgt_len, tgt_len), &self.device)?;
+        let mask = if seqlen_offset > 0 {
+            let mask0 = Tensor::zeros((tgt_len, seqlen_offset), DType::F32, &self.device)?;
+            Tensor::cat(&[&mask0, &mask], D::Minus1)?
+        } else {
+            mask
+        };
+        mask.expand((b_size, 1, tgt_len, tgt_len + seqlen_offset))?
+            .to_dtype(self.dtype)
+    }
+
+    pub fn forward(&mut self, input_ids: &Tensor, seqlen_offset: usize) -> Result<Tensor> {
+        let (b_size, seq_len) = input_ids.dims2()?;
+        let attention_mask = if seq_len <= 1 {
+            None
+        } else {
+            let mask = self.prepare_decoder_attention_mask(b_size, seq_len, seqlen_offset)?;
+            Some(mask)
+        };
+        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() {
+            xs = layer.forward(&xs, attention_mask.as_ref(), seqlen_offset)?
+        }
+        let logits = xs
+            .narrow(1, seq_len - 1, 1)?
+            .apply(&self.norm)?
+            .apply(&self.lm_head)?;
+        let logits = match self.final_logit_softcapping {
+            None => logits,
+            Some(sc) => ((logits / sc)?.tanh()? * sc)?,
+        };
+
+        Ok(logits)
+    }
+
+    pub fn clear_kv_cache(&mut self) {
+        for layer in self.layers.iter_mut() {
+            layer.clear_kv_cache()
+        }
+    }
+}

candle-transformers/src/models/mod.rs

@@ -43,6 +43,7 @@ pub mod fastvit;
 pub mod flux;
 pub mod gemma;
 pub mod gemma2;
+pub mod gemma3;
 pub mod glm4;
 pub mod granite;
 pub mod helium;