Add the StarCoder2 model. (#1779)

* Add the StarCoder2 model.

* Add the example code and get things to work.

* And also tweak the readme.

README.md

@@ -76,7 +76,8 @@ We also provide a some command line based examples using state of the art models
 - [Mixtral8x7b-v0.1](./candle-examples/examples/mixtral/): a sparse mixture of
   experts 8x7b general LLM with better performance than a Llama 2 70B model with
   much faster inference.
-- [StarCoder](./candle-examples/examples/bigcode/): LLM specialized to code generation.
+- [StarCoder](./candle-examples/examples/bigcode/) and
+  [StarCoder2](./candle-examples/examples/starcoder2/): LLM specialized to code generation.
 - [Qwen1.5](./candle-examples/examples/qwen/): Bilingual (English/Chinese) LLMs.
 - [RWKV v5](./candle-examples/examples/rwkv/): An RNN with transformer level LLM
   performance.
@@ -191,7 +192,7 @@ If you have an addition to this list, please submit a pull request.
     - Language Models.
         - LLaMA v1 and v2 with variants such as SOLAR-10.7B.
         - Falcon.
-        - StarCoder.
+        - StarCoder, StarCoder2.
         - Phi 1, 1.5, and 2.
         - Mamba, Minimal Mamba
         - Gemma 2b and 7b.

candle-examples/examples/starcoder2/main.rs

@@ -0,0 +1,253 @@
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+
+use anyhow::{Error as E, Result};
+use clap::Parser;
+
+use candle_transformers::models::starcoder2::Model;
+
+use candle::{DType, Device, Tensor};
+use candle_examples::token_output_stream::TokenOutputStream;
+use candle_nn::VarBuilder;
+use candle_transformers::generation::LogitsProcessor;
+use hf_hub::{api::sync::Api, Repo, RepoType};
+use tokenizers::Tokenizer;
+
+struct TextGeneration {
+    model: Model,
+    device: Device,
+    tokenizer: TokenOutputStream,
+    logits_processor: LogitsProcessor,
+    repeat_penalty: f32,
+    repeat_last_n: usize,
+}
+
+impl TextGeneration {
+    #[allow(clippy::too_many_arguments)]
+    fn new(
+        model: Model,
+        tokenizer: Tokenizer,
+        seed: u64,
+        temp: Option<f64>,
+        top_p: Option<f64>,
+        repeat_penalty: f32,
+        repeat_last_n: usize,
+        device: &Device,
+    ) -> Self {
+        let logits_processor = LogitsProcessor::new(seed, temp, top_p);
+        Self {
+            model,
+            tokenizer: TokenOutputStream::new(tokenizer),
+            logits_processor,
+            repeat_penalty,
+            repeat_last_n,
+            device: device.clone(),
+        }
+    }
+
+    fn run(&mut self, prompt: &str, sample_len: usize) -> Result<()> {
+        use std::io::Write;
+        self.tokenizer.clear();
+        let mut tokens = self
+            .tokenizer
+            .tokenizer()
+            .encode(prompt, true)
+            .map_err(E::msg)?
+            .get_ids()
+            .to_vec();
+        for &t in tokens.iter() {
+            if let Some(t) = self.tokenizer.next_token(t)? {
+                print!("{t}")
+            }
+        }
+        std::io::stdout().flush()?;
+
+        let mut generated_tokens = 0usize;
+        let eos_token = match self.tokenizer.get_token("<|endoftext|>") {
+            Some(token) => token,
+            None => anyhow::bail!("cannot find the <|endoftext|> token"),
+        };
+        let start_gen = std::time::Instant::now();
+        for index in 0..sample_len {
+            let context_size = if index > 0 { 1 } else { tokens.len() };
+            let start_pos = tokens.len().saturating_sub(context_size);
+            let ctxt = &tokens[start_pos..];
+            let input = Tensor::new(ctxt, &self.device)?.unsqueeze(0)?;
+            let logits = self.model.forward(&input, start_pos)?;
+            let logits = logits.squeeze(0)?.squeeze(0)?.to_dtype(DType::F32)?;
+            let logits = if self.repeat_penalty == 1. {
+                logits
+            } else {
+                let start_at = tokens.len().saturating_sub(self.repeat_last_n);
+                candle_transformers::utils::apply_repeat_penalty(
+                    &logits,
+                    self.repeat_penalty,
+                    &tokens[start_at..],
+                )?
+            };
+
+            let next_token = self.logits_processor.sample(&logits)?;
+            tokens.push(next_token);
+            generated_tokens += 1;
+            if next_token == eos_token {
+                break;
+            }
+            if let Some(t) = self.tokenizer.next_token(next_token)? {
+                print!("{t}");
+                std::io::stdout().flush()?;
+            }
+        }
+        let dt = start_gen.elapsed();
+        if let Some(rest) = self.tokenizer.decode_rest().map_err(E::msg)? {
+            print!("{rest}");
+        }
+        std::io::stdout().flush()?;
+        println!(
+            "\n{generated_tokens} tokens generated ({:.2} token/s)",
+            generated_tokens as f64 / dt.as_secs_f64(),
+        );
+        Ok(())
+    }
+}
+
+#[derive(Parser, Debug)]
+#[command(author, version, about, long_about = None)]
+struct Args {
+    /// Run on CPU rather than on GPU.
+    #[arg(long)]
+    cpu: bool,
+
+    /// Enable tracing (generates a trace-timestamp.json file).
+    #[arg(long)]
+    tracing: bool,
+
+    #[arg(long)]
+    use_flash_attn: bool,
+
+    #[arg(long)]
+    prompt: String,
+
+    /// The temperature used to generate samples.
+    #[arg(long)]
+    temperature: Option<f64>,
+
+    /// Nucleus sampling probability cutoff.
+    #[arg(long)]
+    top_p: Option<f64>,
+
+    /// The seed to use when generating random samples.
+    #[arg(long, default_value_t = 299792458)]
+    seed: u64,
+
+    /// The length of the sample to generate (in tokens).
+    #[arg(long, short = 'n', default_value_t = 10000)]
+    sample_len: usize,
+
+    #[arg(long)]
+    model_id: Option<String>,
+
+    #[arg(long, default_value = "main")]
+    revision: String,
+
+    #[arg(long)]
+    config_file: Option<String>,
+
+    #[arg(long)]
+    tokenizer_file: Option<String>,
+
+    #[arg(long)]
+    weight_files: Option<String>,
+
+    /// Penalty to be applied for repeating tokens, 1. means no penalty.
+    #[arg(long, default_value_t = 1.1)]
+    repeat_penalty: f32,
+
+    /// The context size to consider for the repeat penalty.
+    #[arg(long, default_value_t = 64)]
+    repeat_last_n: usize,
+}
+
+fn main() -> Result<()> {
+    use tracing_chrome::ChromeLayerBuilder;
+    use tracing_subscriber::prelude::*;
+
+    let args = Args::parse();
+    let _guard = if args.tracing {
+        let (chrome_layer, guard) = ChromeLayerBuilder::new().build();
+        tracing_subscriber::registry().with(chrome_layer).init();
+        Some(guard)
+    } else {
+        None
+    };
+    println!(
+        "avx: {}, neon: {}, simd128: {}, f16c: {}",
+        candle::utils::with_avx(),
+        candle::utils::with_neon(),
+        candle::utils::with_simd128(),
+        candle::utils::with_f16c()
+    );
+    println!(
+        "temp: {:.2} repeat-penalty: {:.2} repeat-last-n: {}",
+        args.temperature.unwrap_or(0.),
+        args.repeat_penalty,
+        args.repeat_last_n
+    );
+
+    let start = std::time::Instant::now();
+    let api = Api::new()?;
+    let model_id = match args.model_id {
+        Some(model_id) => model_id,
+        None => "bigcode/starcoder2-3b".to_string(),
+    };
+    let repo = api.repo(Repo::with_revision(
+        model_id,
+        RepoType::Model,
+        args.revision,
+    ));
+    let config_file = match args.config_file {
+        Some(file) => std::path::PathBuf::from(file),
+        None => repo.get("config.json")?,
+    };
+    let tokenizer_file = match args.tokenizer_file {
+        Some(file) => std::path::PathBuf::from(file),
+        None => repo.get("tokenizer.json")?,
+    };
+    let filenames = match args.weight_files {
+        Some(files) => files
+            .split(',')
+            .map(std::path::PathBuf::from)
+            .collect::<Vec<_>>(),
+        None => vec![repo.get("model.safetensors")?],
+    };
+    println!("retrieved the files in {:?}", start.elapsed());
+    let tokenizer = Tokenizer::from_file(tokenizer_file).map_err(E::msg)?;
+
+    let start = std::time::Instant::now();
+    let config = serde_json::from_reader(std::fs::File::open(config_file)?)?;
+    let device = candle_examples::device(args.cpu)?;
+    let dtype = if device.is_cuda() {
+        DType::BF16
+    } else {
+        DType::F32
+    };
+    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, dtype, &device)? };
+    let model = Model::new(&config, vb)?;
+
+    println!("loaded the model in {:?}", start.elapsed());
+
+    let mut pipeline = TextGeneration::new(
+        model,
+        tokenizer,
+        args.seed,
+        args.temperature,
+        args.top_p,
+        args.repeat_penalty,
+        args.repeat_last_n,
+        &device,
+    );
+    pipeline.run(&args.prompt, args.sample_len)?;
+    Ok(())
+}

candle-examples/src/token_output_stream.rs

@@ -40,7 +40,7 @@ impl TokenOutputStream {
         };
         self.tokens.push(token);
         let text = self.decode(&self.tokens[self.prev_index..])?;
-        if text.len() > prev_text.len() && text.chars().last().unwrap().is_alphabetic() {
+        if text.len() > prev_text.len() && text.chars().last().unwrap().is_alphanumeric() {
             let text = text.split_at(prev_text.len());
             self.prev_index = self.current_index;
             self.current_index = self.tokens.len();

candle-nn/src/activation.rs

@@ -5,6 +5,7 @@ use serde::Deserialize;
 #[serde(rename_all = "lowercase")]
 pub enum Activation {
     #[default]
+    #[serde(alias = "gelu")]
     Gelu,
     #[serde(alias = "gelu_new")]
     NewGelu,
@@ -19,6 +20,8 @@ pub enum Activation {
     HardSwish,
     Elu(f64),
     LeakyRelu(f64),
+    #[serde(alias = "gelu_pytorch_tanh")]
+    GeluPytorchTanh,
 }
 
 impl super::Module for Activation {
@@ -38,6 +41,7 @@ impl super::Module for Activation {
             Self::HardSwish => xs * crate::ops::hard_sigmoid(xs)?,
             &Self::Elu(alpha) => xs.elu(alpha),
             &Self::LeakyRelu(negative_slope) => crate::ops::leaky_relu(xs, negative_slope),
+            Self::GeluPytorchTanh => xs.gelu(),
         }
     }
 }

candle-transformers/src/models/mod.rs

@@ -41,6 +41,7 @@ pub mod rwkv_v5;
 pub mod segment_anything;
 pub mod stable_diffusion;
 pub mod stable_lm;
+pub mod starcoder2;
 pub mod t5;
 pub mod trocr;
 pub mod vgg;

candle-transformers/src/models/starcoder2.rs

@@ -0,0 +1,347 @@
+#![allow(unused)]
+use candle::{DType, Device, Module, Result, Tensor, D};
+use candle_nn::{layer_norm, linear_b, LayerNorm, Linear, VarBuilder};
+use std::sync::Arc;
+
+#[derive(Debug, Clone, serde::Deserialize)]
+pub struct Config {
+    vocab_size: usize,
+    hidden_size: usize,
+    intermediate_size: usize,
+    num_hidden_layers: usize,
+    num_attention_heads: usize,
+    num_key_value_heads: usize,
+    hidden_act: candle_nn::Activation,
+    max_position_embeddings: usize,
+    norm_epsilon: f64,
+    rope_theta: f64,
+    use_bias: bool,
+    sliding_window: Option<usize>,
+}
+
+#[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.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)?;
+        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))
+    }
+}
+
+#[derive(Debug, Clone)]
+#[allow(clippy::upper_case_acronyms)]
+struct MLP {
+    c_fc: Linear,
+    c_proj: Linear,
+    act: candle_nn::Activation,
+}
+
+impl MLP {
+    fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
+        let (h_size, i_size) = (cfg.hidden_size, cfg.intermediate_size);
+        let c_fc = linear_b(h_size, i_size, cfg.use_bias, vb.pp("c_fc"))?;
+        let c_proj = linear_b(i_size, h_size, cfg.use_bias, vb.pp("c_proj"))?;
+        Ok(Self {
+            c_fc,
+            c_proj,
+            act: cfg.hidden_act,
+        })
+    }
+}
+
+impl Module for MLP {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        xs.apply(&self.c_fc)?.apply(&self.act)?.apply(&self.c_proj)
+    }
+}
+
+#[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)>,
+}
+
+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 b = cfg.use_bias;
+        let q_proj = linear_b(hidden_sz, num_heads * head_dim, b, vb.pp("q_proj"))?;
+        let k_proj = linear_b(hidden_sz, num_kv_heads * head_dim, b, vb.pp("k_proj"))?;
+        let v_proj = linear_b(hidden_sz, num_kv_heads * head_dim, b, vb.pp("v_proj"))?;
+        let o_proj = linear_b(num_heads * head_dim, hidden_sz, b, 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,
+        })
+    }
+
+    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 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)?;
+        let attn_output = attn_weights.matmul(&value_states)?;
+        attn_output
+            .transpose(1, 2)?
+            .reshape((b_sz, q_len, self.hidden_size))?
+            .apply(&self.o_proj)
+    }
+
+    fn clear_kv_cache(&mut self) {
+        self.kv_cache = None
+    }
+}
+
+#[derive(Debug, Clone)]
+struct DecoderLayer {
+    self_attn: Attention,
+    mlp: MLP,
+    input_layernorm: LayerNorm,
+    post_attention_layernorm: LayerNorm,
+}
+
+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 mlp = MLP::new(cfg, vb.pp("mlp"))?;
+        let input_layernorm =
+            layer_norm(cfg.hidden_size, cfg.norm_epsilon, vb.pp("input_layernorm"))?;
+        let post_attention_layernorm = layer_norm(
+            cfg.hidden_size,
+            cfg.norm_epsilon,
+            vb.pp("post_attention_layernorm"),
+        )?;
+        Ok(Self {
+            self_attn,
+            mlp,
+            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.mlp)?;
+        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: LayerNorm,
+    lm_head: Linear,
+    sliding_window: Option<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 = layer_norm(cfg.hidden_size, cfg.norm_epsilon, vb_m.pp("norm"))?;
+        let lm_head = candle_nn::Linear::new(embed_tokens.embeddings().clone(), None);
+        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> {
+        let sliding_window = self.sliding_window.unwrap_or(tgt_len + 42);
+        let mask: Vec<_> = (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 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)
+    }
+
+    pub fn clear_kv_cache(&mut self) {
+        for layer in self.layers.iter_mut() {
+            layer.clear_kv_cache()
+        }
+    }
+}