Add the Mixtral model. (#1437)

* Add the Mixtral model.

* Add more of the mixtral layers.

* Add the final layers for mixtral.

* Sketch the expert selection.

* Add some expert routing logic.

* Hopefully finish the routing logic for mixtral.

* Add the mixtral example.

* Fix the weight filenames.

* Bugfix.

* Another fix.

* Yet another fix + remove the unused pragma.

* Shape fix.

* Add a readme.

2 files changed, 500 insertions, 0 deletions

diff --git a/candle-transformers/src/models/mixtral.rs b/candle-transformers/src/models/mixtral.rs
new file mode 100644
index 00000000..ede74d3f
--- /dev/null
+++ b/candle-transformers/src/models/mixtral.rs
@@ -0,0 +1,499 @@
+use crate::models::with_tracing::{linear_no_bias, Linear};
+/// Mixtral Model
+/// https://github.com/huggingface/transformers/blob/main/src/transformers/models/mixtral/modeling_mixtral.py
+/// https://mistral.ai/news/mixtral-of-experts/
+use candle::{DType, Device, Module, Result, Tensor, D};
+use candle_nn::{Activation, VarBuilder};
+use serde::Deserialize;
+use std::sync::Arc;
+
+/// https://github.com/huggingface/transformers/blob/1a585c1222a56bcaecc070966d558d4a9d862e83/src/transformers/models/mixtral/configuration_mixtral.py#L113
+#[derive(Debug, Clone, PartialEq, Deserialize)]
+pub struct Config {
+    pub(crate) vocab_size: usize,
+    pub(crate) hidden_size: usize,
+    pub(crate) intermediate_size: usize,
+    pub(crate) num_hidden_layers: usize,
+    pub(crate) num_attention_heads: usize,
+    pub(crate) num_key_value_heads: usize,
+    pub(crate) hidden_act: Activation,
+    pub(crate) max_position_embeddings: usize,
+    pub(crate) rms_norm_eps: f64,
+    pub(crate) rope_theta: f64,
+    pub(crate) sliding_window: usize,
+    pub(crate) num_experts_per_tok: usize,
+    pub(crate) num_local_experts: usize,
+    pub(crate) use_flash_attn: bool,
+}
+
+impl Config {
+    /// https://huggingface.co/mistralai/Mixtral-8x7B-v0.1/blob/main/config.json
+    pub fn v0_1_8x7b(use_flash_attn: bool) -> Self {
+        Self {
+            vocab_size: 32000,
+            hidden_size: 4096,
+            intermediate_size: 14336,
+            num_hidden_layers: 32,
+            num_attention_heads: 32,
+            num_key_value_heads: 8,
+            hidden_act: Activation::Silu,
+            max_position_embeddings: 32768,
+            rms_norm_eps: 1e-5,
+            rope_theta: 1e6,
+            sliding_window: 4096,
+            num_experts_per_tok: 2,
+            num_local_experts: 8,
+            use_flash_attn,
+        }
+    }
+}
+
+#[derive(Debug, Clone)]
+struct RmsNorm {
+    inner: candle_nn::RmsNorm,
+    span: tracing::Span,
+}
+
+impl RmsNorm {
+    fn new(size: usize, eps: f64, vb: VarBuilder) -> Result<Self> {
+        let span = tracing::span!(tracing::Level::TRACE, "rms-norm");
+        let inner = candle_nn::rms_norm(size, eps, vb)?;
+        Ok(Self { inner, span })
+    }
+}
+
+impl Module for RmsNorm {
+    fn forward(&self, x: &Tensor) -> Result<Tensor> {
+        let _enter = self.span.enter();
+        self.inner.forward(x)
+    }
+}
+
+#[derive(Debug, Clone)]
+struct RotaryEmbedding {
+    sin: Tensor,
+    cos: Tensor,
+}
+
+fn rotate_half(xs: &Tensor) -> Result<Tensor> {
+    let last_dim = xs.dim(D::Minus1)?;
+    let xs1 = xs.narrow(D::Minus1, 0, last_dim / 2)?;
+    let xs2 = xs.narrow(D::Minus1, last_dim / 2, last_dim - last_dim / 2)?;
+    Tensor::cat(&[&xs2.neg()?, &xs1], D::Minus1)
+}
+
+impl RotaryEmbedding {
+    fn new(dtype: DType, cfg: &Config, dev: &Device) -> Result<Self> {
+        let dim = cfg.hidden_size / cfg.num_attention_heads;
+        let max_seq_len = cfg.max_position_embeddings;
+        let inv_freq: Vec<_> = (0..dim)
+            .step_by(2)
+            .map(|i| 1f32 / (cfg.rope_theta as f32).powf(i as f32 / dim 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)?;
+        let freqs = Tensor::cat(&[&freqs, &freqs], D::Minus1)?;
+        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 cos = cos.unsqueeze(0)?.unsqueeze(0)?; // (1, 1, seq_len, dim)
+        let sin = sin.unsqueeze(0)?.unsqueeze(0)?; // (1, 1, seq_len, dim)
+        let q_embed = (q.broadcast_mul(&cos)? + rotate_half(q)?.broadcast_mul(&sin))?;
+        let k_embed = (k.broadcast_mul(&cos)? + rotate_half(k)?.broadcast_mul(&sin))?;
+        Ok((q_embed, k_embed))
+    }
+}
+
+#[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 Attention {
+    q_proj: Linear,
+    k_proj: Linear,
+    v_proj: Linear,
+    o_proj: Linear,
+    num_heads: usize,
+    num_kv_heads: usize,
+    num_kv_groups: usize,
+    head_dim: usize,
+    hidden_size: usize,
+    rotary_emb: Arc<RotaryEmbedding>,
+    kv_cache: Option<(Tensor, Tensor)>,
+    use_flash_attn: bool,
+}
+
+impl Attention {
+    fn new(rotary_emb: Arc<RotaryEmbedding>, 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 = hidden_sz / num_heads;
+        let q_proj = linear_no_bias(hidden_sz, num_heads * head_dim, vb.pp("q_proj"))?;
+        let k_proj = linear_no_bias(hidden_sz, num_kv_heads * head_dim, vb.pp("k_proj"))?;
+        let v_proj = linear_no_bias(hidden_sz, num_kv_heads * head_dim, vb.pp("v_proj"))?;
+        let o_proj = linear_no_bias(num_heads * head_dim, hidden_sz, vb.pp("o_proj"))?;
+        Ok(Self {
+            q_proj,
+            k_proj,
+            v_proj,
+            o_proj,
+            num_heads,
+            num_kv_heads,
+            num_kv_groups,
+            head_dim,
+            hidden_size: hidden_sz,
+            rotary_emb,
+            kv_cache: None,
+            use_flash_attn: cfg.use_flash_attn,
+        })
+    }
+
+    fn repeat_kv(&self, xs: Tensor) -> Result<Tensor> {
+        let n_rep = self.num_kv_groups;
+        if n_rep == 1 {
+            Ok(xs)
+        } else {
+            let (b_sz, num_kv_heads, seq_len, head_dim) = xs.dims4()?;
+            xs.unsqueeze(2)?
+                .expand((b_sz, num_kv_heads, n_rep, seq_len, head_dim))?
+                .reshape((b_sz, num_kv_heads * n_rep, seq_len, head_dim))
+        }
+    }
+
+    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, key_states) =
+            self.rotary_emb
+                .apply_rotary_emb_qkv(&query_states, &key_states, seqlen_offset)?;
+
+        let (key_states, value_states) = match &self.kv_cache {
+            None => (key_states, value_states),
+            Some((prev_k, prev_v)) => {
+                let key_states = Tensor::cat(&[prev_k, &key_states], 2)?;
+                let value_states = Tensor::cat(&[prev_v, &value_states], 2)?;
+                (key_states, value_states)
+            }
+        };
+        self.kv_cache = Some((key_states.clone(), value_states.clone()));
+
+        let key_states = self.repeat_kv(key_states)?;
+        let value_states = self.repeat_kv(value_states)?;
+
+        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 softmax_scale = 1f32 / (self.head_dim as f32).sqrt();
+            flash_attn(&q, &k, &v, softmax_scale, q_len > 1)?.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 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, self.hidden_size))?
+            .apply(&self.o_proj)
+    }
+}
+
+#[derive(Debug, Clone)]
+struct BlockSparseTop2MLP {
+    w1: Linear,
+    w2: Linear,
+    w3: Linear,
+    act_fn: Activation,
+}
+
+impl BlockSparseTop2MLP {
+    fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
+        let hidden_sz = cfg.hidden_size;
+        let intermediate_sz = cfg.intermediate_size;
+        let w1 = linear_no_bias(hidden_sz, intermediate_sz, vb.pp("w1"))?;
+        let w2 = linear_no_bias(intermediate_sz, hidden_sz, vb.pp("w2"))?;
+        let w3 = linear_no_bias(hidden_sz, intermediate_sz, vb.pp("w3"))?;
+        Ok(Self {
+            w1,
+            w2,
+            w3,
+            act_fn: cfg.hidden_act,
+        })
+    }
+}
+
+impl Module for BlockSparseTop2MLP {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        let lhs = xs.apply(&self.w1)?.apply(&self.act_fn)?;
+        let rhs = xs.apply(&self.w3)?;
+        (lhs * rhs)?.apply(&self.w2)
+    }
+}
+
+#[derive(Debug, Clone)]
+struct SparseMoeBlock {
+    gate: Linear,
+    experts: Vec<BlockSparseTop2MLP>,
+    num_experts_per_tok: usize,
+}
+
+impl SparseMoeBlock {
+    fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
+        let gate = linear_no_bias(cfg.hidden_size, cfg.num_local_experts, vb.pp("gate"))?;
+        let mut experts = Vec::with_capacity(cfg.num_local_experts);
+        let vb = vb.pp("experts");
+        for idx in 0..cfg.num_local_experts {
+            let expert = BlockSparseTop2MLP::new(cfg, vb.pp(idx))?;
+            experts.push(expert)
+        }
+        Ok(SparseMoeBlock {
+            gate,
+            experts,
+            num_experts_per_tok: cfg.num_experts_per_tok,
+        })
+    }
+}
+
+impl Module for SparseMoeBlock {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        let (b_size, seq_len, hidden_dim) = xs.dims3()?;
+        let xs = xs.reshape(((), hidden_dim))?;
+        let router_logits = xs.apply(&self.gate)?;
+        let routing_weights = candle_nn::ops::softmax_last_dim(&router_logits)?;
+
+        // In order to extract topk, we extract the data from the tensor and manipulate it
+        // directly. Maybe we will want to use some custom ops instead at some point.
+        let routing_weights = routing_weights.to_dtype(DType::F32)?.to_vec2::<f32>()?;
+
+        // routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+        // top_x contains the row indexes to evaluate for each expert.
+        let mut top_x = vec![vec![]; self.experts.len()];
+        let mut selected_rws = vec![vec![]; self.experts.len()];
+        for (row_idx, rw) in routing_weights.iter().enumerate() {
+            let mut dst = (0..rw.len() as u32).collect::<Vec<u32>>();
+            dst.sort_by(|&i, &j| rw[j as usize].total_cmp(&rw[i as usize]));
+            let mut sum_routing_weights = 0f32;
+            for &expert_idx in dst.iter().take(self.num_experts_per_tok) {
+                let expert_idx = expert_idx as usize;
+                let routing_weight = rw[expert_idx];
+                sum_routing_weights += routing_weight;
+                top_x[expert_idx].push(row_idx as u32);
+            }
+            for &expert_idx in dst.iter().take(self.num_experts_per_tok) {
+                let expert_idx = expert_idx as usize;
+                let routing_weight = rw[expert_idx];
+                selected_rws[expert_idx].push(routing_weight / sum_routing_weights)
+            }
+        }
+
+        // routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+        // expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
+
+        let mut ys = xs.zeros_like()?;
+        for (expert_idx, expert_layer) in self.experts.iter().enumerate() {
+            let top_x = &top_x[expert_idx];
+            if top_x.is_empty() {
+                continue;
+            }
+            let top_x = Tensor::new(top_x.as_slice(), xs.device())?;
+            let selected_rws =
+                Tensor::new(selected_rws[expert_idx].as_slice(), xs.device())?.reshape(((), 1))?;
+            // Index the correct hidden states and compute the expert hidden state for
+            // the current expert. We need to make sure to multiply the output hidden
+            // states by `routing_weights` on the corresponding tokens (top-1 and top-2)
+            let current_state = xs.index_select(&top_x, 0)?.reshape(((), hidden_dim))?;
+            // current_hidden_states = expert_layer(current_state, routing_weights[top_x_list, idx_list, None])
+            let current_hidden_states = expert_layer.forward(&current_state)?;
+            let current_hidden_states = current_hidden_states.broadcast_mul(&selected_rws)?;
+            ys = ys.index_add(&top_x, &current_hidden_states, 0)?;
+        }
+
+        let ys = ys.reshape((b_size, seq_len, hidden_dim))?;
+        Ok(ys)
+    }
+}
+
+#[derive(Debug, Clone)]
+struct DecoderLayer {
+    self_attn: Attention,
+    block_sparse_moe: SparseMoeBlock,
+    input_layernorm: RmsNorm,
+    post_attention_layernorm: RmsNorm,
+}
+
+impl DecoderLayer {
+    fn new(rotary_emb: Arc<RotaryEmbedding>, cfg: &Config, vb: VarBuilder) -> Result<Self> {
+        let self_attn = Attention::new(rotary_emb, cfg, vb.pp("self_attn"))?;
+        let block_sparse_moe = SparseMoeBlock::new(cfg, vb.pp("block_sparse_moe"))?;
+        let input_layernorm =
+            RmsNorm::new(cfg.hidden_size, cfg.rms_norm_eps, vb.pp("input_layernorm"))?;
+        let post_attention_layernorm = RmsNorm::new(
+            cfg.hidden_size,
+            cfg.rms_norm_eps,
+            vb.pp("post_attention_layernorm"),
+        )?;
+        Ok(Self {
+            self_attn,
+            block_sparse_moe,
+            input_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 + residual)?;
+        let residual = &xs;
+        let xs = xs
+            .apply(&self.post_attention_layernorm)?
+            .apply(&self.block_sparse_moe)?;
+        residual + xs
+    }
+}
+
+#[derive(Debug, Clone)]
+pub struct Model {
+    embed_tokens: candle_nn::Embedding,
+    layers: Vec<DecoderLayer>,
+    norm: RmsNorm,
+    lm_head: Linear,
+    sliding_window: usize,
+    device: Device,
+    dtype: DType,
+}
+
+impl Model {
+    pub fn new(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 layer = DecoderLayer::new(rotary_emb.clone(), 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_no_bias(cfg.hidden_size, cfg.vocab_size, vb.pp("lm_head"))?;
+        Ok(Self {
+            embed_tokens,
+            layers,
+            norm,
+            lm_head,
+            sliding_window: cfg.sliding_window,
+            device: vb.device().clone(),
+            dtype: vb.dtype(),
+        })
+    }
+
+    fn prepare_decoder_attention_mask(
+        &self,
+        b_size: usize,
+        tgt_len: usize,
+        seqlen_offset: usize,
+    ) -> Result<Tensor> {
+        // Sliding window mask?
+        let mask: Vec<_> = (0..tgt_len)
+            .flat_map(|i| {
+                (0..tgt_len).map(move |j| {
+                    if i < j || j + self.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 mut xs = self.embed_tokens.forward(input_ids)?;
+        for layer in self.layers.iter_mut() {
+            xs = layer.forward(&xs, attention_mask.as_ref(), seqlen_offset)?
+        }
+        xs.narrow(1, seq_len - 1, 1)?
+            .apply(&self.norm)?
+            .apply(&self.lm_head)
+    }
+}
diff --git a/candle-transformers/src/models/mod.rs b/candle-transformers/src/models/mod.rs
index a9a56673..94a3bd5b 100644
--- a/candle-transformers/src/models/mod.rs
+++ b/candle-transformers/src/models/mod.rs
@@ -14,6 +14,7 @@ pub mod llama2_c_weights;
 pub mod marian;
 pub mod mistral;
 pub mod mixformer;
+pub mod mixtral;
 pub mod mpt;
 pub mod persimmon;
 pub mod quantized_blip;