Fixes for clippy 1.86.

candle-examples/examples/csm/main.rs

@@ -0,0 +1,221 @@
+#[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::csm::{Config, Model};
+
+use candle::{DType, IndexOp, Tensor};
+use candle_nn::VarBuilder;
+use hf_hub::{api::sync::Api, Repo, RepoType};
+use tokenizers::Tokenizer;
+
+#[derive(Clone, Debug, Copy, PartialEq, Eq, clap::ValueEnum)]
+enum Which {
+    #[value(name = "1b")]
+    Csm1b,
+}
+
+#[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, default_value = "[0]Hey how are you doing?")]
+    prompt: String,
+
+    /// The temperature used to generate samples.
+    #[arg(long, default_value_t = 0.7)]
+    temperature: f64,
+
+    /// Nucleus sampling probability cutoff.
+    #[arg(long)]
+    top_p: Option<f64>,
+
+    /// Only sample among the top K samples.
+    #[arg(long)]
+    top_k: Option<usize>,
+
+    /// 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,
+
+    /// The model size to use.
+    #[arg(long, default_value = "1b")]
+    which: Which,
+
+    #[arg(long)]
+    model_id: Option<String>,
+
+    #[arg(long, default_value = "main")]
+    revision: String,
+
+    #[arg(long)]
+    tokenizer: Option<String>,
+
+    #[arg(long)]
+    config: Option<String>,
+
+    #[arg(long)]
+    weights: Option<String>,
+
+    /// The mimi model weight file, in safetensor format.
+    #[arg(long)]
+    mimi_weights: 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, 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 => {
+            let name = match args.which {
+                Which::Csm1b => "sesame/csm-1b",
+            };
+            name.to_string()
+        }
+    };
+    let repo = api.repo(Repo::with_revision(
+        model_id,
+        RepoType::Model,
+        args.revision,
+    ));
+    let filenames = match args.weights {
+        Some(files) => files
+            .split(',')
+            .map(std::path::PathBuf::from)
+            .collect::<Vec<_>>(),
+        None => vec![repo.get("model.safetensors")?],
+    };
+    let tokenizer_filename = match args.tokenizer {
+        Some(file) => std::path::PathBuf::from(file),
+        None => api
+            .model("meta-llama/Llama-3.2-1B".to_string())
+            .get("tokenizer.json")?,
+    };
+    let mimi_filename = match args.mimi_weights {
+        Some(model) => std::path::PathBuf::from(model),
+        None => Api::new()?
+            .model("kyutai/mimi".to_string())
+            .get("model.safetensors")?,
+    };
+    println!("retrieved the files in {:?}", start.elapsed());
+    let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
+
+    let start = std::time::Instant::now();
+    let config: Config = match args.config {
+        Some(config_file) => serde_json::from_slice(&std::fs::read(config_file)?)?,
+        None => {
+            let config_file = repo.get("config.json")?;
+            serde_json::from_slice(&std::fs::read(config_file)?)?
+        }
+    };
+    let device = candle_examples::device(args.cpu)?;
+    let (mut model, device) = {
+        let dtype = DType::F32;
+        let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, dtype, &device)? };
+        let model = Model::new(&config, vb)?;
+        (model, device)
+    };
+    let mut mimi_model = {
+        use candle_transformers::models::mimi;
+        let vb =
+            unsafe { VarBuilder::from_mmaped_safetensors(&[mimi_filename], DType::F32, &device)? };
+        let config = mimi::Config::v0_1(Some(32));
+        mimi::Model::new(config, vb)?
+    };
+    let cb = config.audio_num_codebooks;
+
+    println!("loaded the model in {:?}", start.elapsed());
+    if args.prompt.ends_with(".safetensors") {
+        let prompt = candle::safetensors::load(args.prompt, &device)?;
+        let mut tokens = prompt
+            .get("tokens")
+            .expect("no tokens in prompt")
+            .to_dtype(DType::U32)?;
+        let mut mask = prompt.get("mask").expect("no mask in prompt").clone();
+        println!("tokens:\n{tokens:?}");
+        println!("mask:\n{mask:?}");
+        let mut lp = candle_transformers::generation::LogitsProcessor::new(42, None, None);
+        let mut const_mask = vec![1u8; cb];
+        const_mask.push(0);
+        let const_mask = Tensor::from_vec(const_mask, (1, 1, cb + 1), &device)?;
+        let mut pos = 0;
+        let mut all_tokens = vec![];
+        for i in 0.. {
+            let mut frame = model.generate_frame(&tokens, &mask, pos, &mut lp)?;
+            pos += tokens.dim(1)?;
+            frame.push(0);
+            if frame.iter().all(|&x| x == 0) {
+                break;
+            }
+            println!("frame {i} {pos}:\n{frame:?}");
+            tokens = Tensor::from_vec(frame, (1, 1, cb + 1), &device)?;
+            all_tokens.push(tokens.clone());
+            mask = const_mask.clone();
+        }
+        let all_tokens = Tensor::cat(&all_tokens, 1)?.narrow(2, 0, cb)?.t()?;
+        println!("all_tokens:\n{all_tokens:?}");
+        let pcm = mimi_model.decode(&all_tokens)?;
+        let pcm = pcm.i(0)?.i(0)?.to_dtype(DType::F32)?;
+        let pcm = candle_examples::audio::normalize_loudness(&pcm, 24_000, true)?;
+        let pcm = pcm.to_vec1::<f32>()?;
+        let mut output = std::fs::File::create("out.wav")?;
+        candle_examples::wav::write_pcm_as_wav(&mut output, &pcm, 24_000)?;
+    } else {
+        let prompt = tokenizer.encode(args.prompt, true).map_err(E::msg)?;
+        println!("{prompt:?}");
+    }
+
+    Ok(())
+}

candle-nn/src/loss.rs

@@ -7,7 +7,7 @@ use candle::{Result, Tensor};
 /// Arguments
 ///
 /// * [inp]: The input tensor of dimensions `N, C` where `N` is the batch size and `C` the number
-///          of categories. This is expected to contain log probabilities.
+///   of categories. This is expected to contain log probabilities.
 /// * [target]: The ground truth labels as a tensor of u32 of dimension `N`.
 ///
 /// The resulting tensor is a scalar containing the average value over the batch.
@@ -34,7 +34,7 @@ pub fn nll(inp: &Tensor, target: &Tensor) -> Result<Tensor> {
 /// Arguments
 ///
 /// * [inp]: The input tensor of dimensions `N, C` where `N` is the batch size and `C` the number
-///          of categories. This is expected to raw logits.
+///   of categories. This is expected to raw logits.
 /// * [target]: The ground truth labels as a tensor of u32 of dimension `N`.
 ///
 /// The resulting tensor is a scalar containing the average value over the batch.
@@ -56,9 +56,9 @@ pub fn mse(inp: &Tensor, target: &Tensor) -> Result<Tensor> {
 /// Arguments
 ///
 /// * [inp]: The input tensor of dimensions `N, C` where `N` is the batch size and `C` the number
-///          of categories. This is expected to raw logits.
+///   of categories. This is expected to raw logits.
 /// * [target]: The ground truth labels as a tensor of u32 of dimension `N, C` where `N` is the batch size and `C` the number
-///          of categories.
+///   of categories.
 ///
 /// The resulting tensor is a scalar containing the average value over the batch.
 pub fn binary_cross_entropy_with_logit(inp: &Tensor, target: &Tensor) -> Result<Tensor> {

candle-transformers/src/models/csm.rs

@@ -0,0 +1,501 @@
+//! Implementation of the Conversational Speech Model (CSM) from Sesame
+//!
+//! See: [CSM](Conversational Speech Model)
+//!
+/// CSM (Conversational Speech Model) is a speech generation model from Sesame that generates RVQ
+/// audio codes from text and audio inputs. The model architecture employs a Llama backbone and a
+/// smaller audio decoder that produces Mimi audio codes.
+///
+use crate::generation::LogitsProcessor;
+use candle::{DType, Device, IndexOp, Module, Result, Tensor, D};
+use candle_nn::{embedding, linear_b, Embedding, Linear, RmsNorm, VarBuilder};
+use std::sync::Arc;
+
+#[derive(serde::Deserialize, Debug, Clone, Copy, PartialEq, Eq)]
+pub enum Flavor {
+    #[serde(rename = "llama-1B")]
+    Llama1B,
+    #[serde(rename = "llama-100M")]
+    Llama100M,
+}
+
+#[derive(serde::Deserialize, Debug, Clone)]
+pub struct Config {
+    pub audio_num_codebooks: usize,
+    pub audio_vocab_size: usize,
+    pub backbone_flavor: Flavor,
+    pub decoder_flavor: Flavor,
+    pub text_vocab_size: usize,
+}
+
+#[allow(unused)]
+#[derive(Debug, Clone)]
+pub struct LlamaConfig {
+    vocab_size: usize,
+    num_layers: usize,
+    num_heads: usize,
+    num_kv_heads: usize,
+    embed_dim: usize,
+    max_seq_len: usize,
+    intermediate_dim: usize,
+    norm_eps: f64,
+    rope_base: f32,
+    scale_factor: usize,
+}
+
+impl LlamaConfig {
+    pub fn from_flavor(flavor: Flavor) -> Self {
+        match flavor {
+            Flavor::Llama1B => Self {
+                vocab_size: 128256,
+                num_layers: 16,
+                num_heads: 32,
+                num_kv_heads: 8,
+                embed_dim: 2048,
+                max_seq_len: 2048,
+                intermediate_dim: 8192,
+                norm_eps: 1e-5,
+                rope_base: 500_000.,
+                scale_factor: 32,
+            },
+            Flavor::Llama100M => Self {
+                vocab_size: 128256,
+                num_layers: 4,
+                num_heads: 8,
+                num_kv_heads: 2,
+                embed_dim: 1024,
+                max_seq_len: 2048,
+                intermediate_dim: 8192,
+                norm_eps: 1e-5,
+                rope_base: 500_000.,
+                scale_factor: 32,
+            },
+        }
+    }
+}
+
+#[derive(Debug, Clone)]
+struct RotaryEmbedding {
+    sin: Tensor,
+    cos: Tensor,
+}
+
+fn calculate_default_inv_freq(cfg: &LlamaConfig) -> Vec<f32> {
+    let head_dim = cfg.embed_dim / cfg.num_heads;
+    (0..head_dim)
+        .step_by(2)
+        .map(|i| 1f32 / cfg.rope_base.powf(i as f32 / head_dim as f32))
+        .collect()
+}
+
+impl RotaryEmbedding {
+    fn new(dtype: DType, cfg: &LlamaConfig, dev: &Device) -> Result<Self> {
+        let low_freq_factor = 1.0;
+        let high_freq_factor = 4.0;
+        let original_max_position_embeddings = 8192;
+        let scale_factor = cfg.scale_factor as f32;
+        let theta = {
+            let low_freq_wavelen = original_max_position_embeddings as f32 / low_freq_factor;
+            let high_freq_wavelen = original_max_position_embeddings as f32 / high_freq_factor;
+
+            calculate_default_inv_freq(cfg)
+                .into_iter()
+                .map(|freq| {
+                    let wavelen = 2. * std::f32::consts::PI / freq;
+                    if wavelen < high_freq_wavelen {
+                        freq
+                    } else if wavelen > low_freq_wavelen {
+                        freq / scale_factor
+                    } else {
+                        let smooth = (original_max_position_embeddings as f32 / wavelen
+                            - low_freq_factor)
+                            / (high_freq_factor - low_freq_factor);
+                        (1. - smooth) * freq / scale_factor + smooth * freq
+                    }
+                })
+                .collect::<Vec<_>>()
+        };
+
+        let theta = Tensor::new(theta, dev)?;
+        let idx_theta = Tensor::arange(0, cfg.max_seq_len as u32, dev)?
+            .to_dtype(DType::F32)?
+            .reshape((cfg.max_seq_len, 1))?
+            .matmul(&theta.reshape((1, theta.elem_count()))?)?;
+        // This is different from the paper, see:
+        // https://github.com/huggingface/transformers/blob/6112b1c6442aaf7affd2b0676a1cd4eee30c45cf/src/transformers/models/llama/modeling_llama.py#L112
+        let cos = idx_theta.cos()?.to_dtype(dtype)?;
+        let sin = idx_theta.sin()?.to_dtype(dtype)?;
+        Ok(Self { cos, sin })
+    }
+
+    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_i(q, &cos, &sin)?;
+        let k_embed = candle_nn::rotary_emb::rope_i(k, &cos, &sin)?;
+        Ok((q_embed, k_embed))
+    }
+}
+fn rms_norm(hidden_size: usize, eps: f64, vb: VarBuilder) -> Result<RmsNorm> {
+    let weight = vb.get((hidden_size,), "scale")?;
+    Ok(RmsNorm::new(weight, eps))
+}
+
+#[derive(Debug, Clone)]
+struct Attention {
+    q_proj: Linear,
+    k_proj: Linear,
+    v_proj: Linear,
+    o_proj: Linear,
+    rotary_emb: Arc<RotaryEmbedding>,
+    kv_cache: Option<(Tensor, Tensor)>,
+    num_heads: usize,
+    head_dim: usize,
+    num_kv_heads: usize,
+    num_kv_groups: usize,
+}
+
+impl Attention {
+    fn new(cfg: &LlamaConfig, rotary_emb: Arc<RotaryEmbedding>, vb: VarBuilder) -> Result<Self> {
+        let head_dim = cfg.embed_dim / cfg.num_heads;
+        let kv_dim = cfg.num_kv_heads * head_dim;
+
+        let q_proj = linear_b(cfg.embed_dim, cfg.embed_dim, false, vb.pp("q_proj"))?;
+        let k_proj = linear_b(cfg.embed_dim, kv_dim, false, vb.pp("k_proj"))?;
+        let v_proj = linear_b(cfg.embed_dim, kv_dim, false, vb.pp("v_proj"))?;
+        let o_proj = linear_b(cfg.embed_dim, cfg.embed_dim, false, vb.pp("output_proj"))?;
+        Ok(Self {
+            q_proj,
+            k_proj,
+            v_proj,
+            o_proj,
+            rotary_emb,
+            kv_cache: None,
+            num_heads: cfg.num_heads,
+            num_kv_heads: cfg.num_kv_heads,
+            num_kv_groups: cfg.num_heads / cfg.num_kv_heads,
+            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)?
+            .contiguous()?;
+        let key_states = key_states
+            .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
+            .transpose(1, 2)?
+            .contiguous()?;
+        let value_states = value_states
+            .reshape((b_sz, q_len, self.num_kv_heads, self.head_dim))?
+            .transpose(1, 2)?
+            .contiguous()?;
+
+        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 = crate::utils::repeat_kv(key_states, self.num_kv_groups)?;
+        let value_states = crate::utils::repeat_kv(value_states, self.num_kv_groups)?;
+
+        let attn_output = {
+            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.num_heads * self.head_dim))?
+            .apply(&self.o_proj)
+    }
+
+    fn clear_kv_cache(&mut self) {
+        self.kv_cache = None
+    }
+}
+
+#[derive(Debug, Clone)]
+struct Mlp {
+    w1: Linear,
+    w2: Linear,
+    w3: Linear,
+}
+
+impl Mlp {
+    fn new(cfg: &LlamaConfig, vb: VarBuilder) -> Result<Self> {
+        let w1 = linear_b(cfg.embed_dim, cfg.intermediate_dim, false, vb.pp("w1"))?;
+        let w2 = linear_b(cfg.intermediate_dim, cfg.embed_dim, false, vb.pp("w2"))?;
+        let w3 = linear_b(cfg.embed_dim, cfg.intermediate_dim, false, vb.pp("w3"))?;
+        Ok(Self { w1, w2, w3 })
+    }
+}
+
+impl Module for Mlp {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        let lhs = xs.apply(&self.w1)?.silu()?;
+        let rhs = xs.apply(&self.w3)?;
+        (lhs * rhs)?.apply(&self.w2)
+    }
+}
+
+#[derive(Debug, Clone)]
+struct Layer {
+    mlp_norm: RmsNorm,
+    sa_norm: RmsNorm,
+    attn: Attention,
+    mlp: Mlp,
+}
+
+impl Layer {
+    fn new(cfg: &LlamaConfig, rotary_emb: Arc<RotaryEmbedding>, vb: VarBuilder) -> Result<Self> {
+        let mlp_norm = rms_norm(cfg.embed_dim, cfg.norm_eps, vb.pp("mlp_norm"))?;
+        let sa_norm = rms_norm(cfg.embed_dim, cfg.norm_eps, vb.pp("sa_norm"))?;
+        let attn = Attention::new(cfg, rotary_emb, vb.pp("attn"))?;
+        let mlp = Mlp::new(cfg, vb.pp("mlp"))?;
+        Ok(Self {
+            mlp_norm,
+            sa_norm,
+            attn,
+            mlp,
+        })
+    }
+
+    fn forward(
+        &mut self,
+        xs: &Tensor,
+        attention_mask: Option<&Tensor>,
+        seqlen_offset: usize,
+    ) -> Result<Tensor> {
+        let residual = xs;
+        let xs = self.sa_norm.forward(xs)?;
+        let xs = self.attn.forward(&xs, attention_mask, seqlen_offset)?;
+        let xs = (xs + residual)?;
+        let residual = &xs;
+        let xs = xs.apply(&self.mlp_norm)?.apply(&self.mlp)?;
+        residual + xs
+    }
+
+    fn clear_kv_cache(&mut self) {
+        self.attn.clear_kv_cache()
+    }
+}
+
+#[derive(Debug, Clone)]
+pub struct LlamaModel {
+    layers: Vec<Layer>,
+    norm: RmsNorm,
+    device: Device,
+    dtype: DType,
+}
+
+impl LlamaModel {
+    pub fn new(cfg: &LlamaConfig, vb: VarBuilder) -> Result<Self> {
+        let rotary_emb = Arc::new(RotaryEmbedding::new(vb.dtype(), cfg, vb.device())?);
+        let mut layers = Vec::with_capacity(cfg.num_layers);
+        let vb_l = vb.pp("layers");
+        for layer_idx in 0..cfg.num_layers {
+            let layer = Layer::new(cfg, rotary_emb.clone(), vb_l.pp(layer_idx))?;
+            layers.push(layer);
+        }
+        let norm = rms_norm(cfg.embed_dim, cfg.norm_eps, vb.pp("norm"))?;
+        Ok(Self {
+            layers,
+            norm,
+            device: vb.device().clone(),
+            dtype: vb.dtype(),
+        })
+    }
+
+    pub fn clear_kv_cache(&mut self) {
+        for layer in self.layers.iter_mut() {
+            layer.clear_kv_cache()
+        }
+    }
+
+    fn prepare_decoder_attention_mask(
+        &self,
+        tgt_len: usize,
+        seqlen_offset: usize,
+    ) -> Result<Tensor> {
+        let mask: Vec<_> = (0..tgt_len)
+            .flat_map(|i| (0..tgt_len).map(move |j| if i < j { 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((1, 1, tgt_len, tgt_len + seqlen_offset))?
+            .to_dtype(self.dtype)
+    }
+
+    pub fn forward(&mut self, xs: &Tensor, seqlen_offset: usize) -> Result<Tensor> {
+        let (_b_size, seq_len, _embed_dim) = xs.dims3()?;
+        let attention_mask = if seq_len <= 1 {
+            None
+        } else {
+            let mask = self.prepare_decoder_attention_mask(seq_len, seqlen_offset)?;
+            Some(mask)
+        };
+        let mut xs = xs.clone();
+        for layer in self.layers.iter_mut() {
+            xs = layer.forward(&xs, attention_mask.as_ref(), seqlen_offset)?;
+        }
+        let ys = xs.narrow(1, seq_len - 1, 1)?.apply(&self.norm)?;
+        Ok(ys)
+    }
+}
+
+#[derive(Debug, Clone)]
+pub struct Model {
+    backbone: LlamaModel,
+    decoder: LlamaModel,
+    codebook0_head: Linear,
+    audio_embeddings: Embedding,
+    text_embeddings: Embedding,
+    projection: Linear,
+    audio_head: Tensor,
+    config: Config,
+}
+
+impl Model {
+    pub fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
+        let backbone_cfg = LlamaConfig::from_flavor(cfg.backbone_flavor);
+        let backbone = LlamaModel::new(&backbone_cfg, vb.pp("backbone"))?;
+        let decoder_cfg = LlamaConfig::from_flavor(cfg.decoder_flavor);
+        let decoder = LlamaModel::new(&decoder_cfg, vb.pp("decoder"))?;
+        let backbone_dim = backbone_cfg.embed_dim;
+        let decoder_dim = decoder_cfg.embed_dim;
+        let audio_embeddings = embedding(
+            cfg.audio_vocab_size * cfg.audio_num_codebooks,
+            backbone_dim,
+            vb.pp("audio_embeddings"),
+        )?;
+        let text_embeddings =
+            embedding(cfg.text_vocab_size, backbone_dim, vb.pp("text_embeddings"))?;
+        let projection = linear_b(backbone_dim, decoder_dim, false, vb.pp("projection"))?;
+        let codebook0_head = linear_b(
+            backbone_dim,
+            cfg.audio_vocab_size,
+            false,
+            vb.pp("codebook0_head"),
+        )?;
+        let audio_head = vb.get(
+            (
+                cfg.audio_num_codebooks - 1,
+                decoder_dim,
+                cfg.audio_vocab_size,
+            ),
+            "audio_head",
+        )?;
+        Ok(Self {
+            backbone,
+            decoder,
+            codebook0_head,
+            audio_embeddings,
+            text_embeddings,
+            projection,
+            audio_head,
+            config: cfg.clone(),
+        })
+    }
+
+    pub fn clear_kv_cache(&mut self) {
+        self.backbone.clear_kv_cache();
+        self.decoder.clear_kv_cache();
+    }
+
+    pub fn generate_frame(
+        &mut self,
+        tokens: &Tensor,
+        tokens_mask: &Tensor,
+        input_pos: usize,
+        lp: &mut LogitsProcessor,
+    ) -> Result<Vec<u32>> {
+        let (b_sz, seq_len, _cb_plus_one) = tokens.dims3()?;
+        let audio_tokens = tokens.narrow(2, 0, self.config.audio_num_codebooks)?;
+        let text_tokens = tokens.narrow(2, self.config.audio_num_codebooks, 1)?;
+        let text_embeds = self.text_embeddings.forward(&text_tokens)?;
+        let arange = (Tensor::arange(
+            0u32,
+            self.config.audio_num_codebooks as u32,
+            &self.decoder.device,
+        )? * self.config.audio_vocab_size as f64)?;
+        let audio_tokens = audio_tokens.broadcast_add(&arange.reshape((1, 1, ()))?)?;
+        let audio_embeds = self.audio_embeddings.forward(&audio_tokens)?.reshape((
+            b_sz,
+            seq_len,
+            self.config.audio_num_codebooks,
+            (),
+        ))?;
+        let embeds = Tensor::cat(&[&audio_embeds, &text_embeds], D::Minus2)?;
+        let embeds = embeds.broadcast_mul(
+            &tokens_mask
+                .to_dtype(self.backbone.dtype)?
+                .unsqueeze(D::Minus1)?,
+        )?;
+        let embeds = embeds.sum(2)?;
+        let h = self.backbone.forward(&embeds, input_pos)?;
+        let c0_logits = h.apply(&self.codebook0_head)?;
+        let c0_sample = lp.sample(&c0_logits.i((0, 0))?)?;
+        let mut all_samples = vec![c0_sample];
+        let c0_sample = Tensor::from_slice(&[c0_sample], (1, 1), &self.decoder.device)?;
+        let c0_embed = self.audio_embeddings.forward(&c0_sample)?;
+        let mut curr_h = Tensor::cat(&[h, c0_embed], 1)?;
+
+        self.decoder.clear_kv_cache();
+        let mut decoder_pos = 0;
+        for i in 1..self.config.audio_num_codebooks {
+            let proj_h = curr_h.apply(&self.projection)?;
+            let decoder_h = self.decoder.forward(&proj_h, decoder_pos)?;
+            decoder_pos += curr_h.dim(1)?;
+            let ci_logits = decoder_h.broadcast_matmul(&self.audio_head.get(i - 1)?)?;
+            let ci_sample = lp.sample(&ci_logits.i((0, 0))?)?;
+            all_samples.push(ci_sample);
+            let ci_sample = Tensor::from_slice(
+                &[ci_sample + (i * self.config.audio_vocab_size) as u32],
+                (1, 1),
+                &self.decoder.device,
+            )?;
+            let ci_embed = self.audio_embeddings.forward(&ci_sample)?;
+            curr_h = ci_embed
+        }
+        Ok(all_samples)
+    }
+}

candle-transformers/src/models/mod.rs

@@ -27,6 +27,7 @@ pub mod codegeex4_9b;
 pub mod colpali;
 pub mod convmixer;
 pub mod convnext;
+pub mod csm;
 pub mod dac;
 pub mod debertav2;
 pub mod deepseek2;