Merge pull request #41 from LaurentMazare/kv-cache

Kv cache

Cargo.toml

@@ -1,8 +1,6 @@
 [workspace]
-
 members = [
     "candle-core",
     "candle-hub",
     "candle-kernels",
 ]
-

candle-core/examples/llama/main.rs

@@ -24,6 +24,7 @@ mod var_store;
 mod weights;
 
 const CONTEXT_SIZE: usize = 512;
+const USE_KV_CACHE: bool = false;
 const START_PROMPT: &str = r"
 EDWARD:
 I wonder how our princely father 'scaped,
@@ -218,13 +219,16 @@ fn masked_fill(on_false: &Tensor, mask: &Tensor, on_true: f32) -> Result<Tensor>
 #[derive(Clone)]
 struct Cache {
     masks: Arc<Mutex<HashMap<usize, Tensor>>>,
+    #[allow(clippy::type_complexity)]
+    kvs: Arc<Mutex<Vec<Option<(Tensor, Tensor)>>>>,
     device: Device,
 }
 
 impl Cache {
-    fn new(device: &Device) -> Self {
+    fn new(config: &Config, device: &Device) -> Self {
         Self {
             masks: Arc::new(Mutex::new(HashMap::new())),
+            kvs: Arc::new(Mutex::new(vec![None; config.n_layer])),
             device: device.clone(),
         }
     }
@@ -249,7 +253,6 @@ struct CausalSelfAttention {
     c_attn: Linear,
     c_proj: Linear,
     n_head: usize,
-    // n_embd: usize,
     cache: Cache,
 }
 
@@ -265,6 +268,11 @@ impl CausalSelfAttention {
 
     fn apply_rotary_emb(&self, x: &Tensor, freqs_cis: &Tensor) -> Result<Tensor> {
         let mut dims = x.dims().to_vec();
+        let freqs_cis = if dims[1] < CONTEXT_SIZE {
+            freqs_cis.narrow(1, CONTEXT_SIZE - dims[1], dims[1])?
+        } else {
+            freqs_cis.clone()
+        };
         let v = dims.pop().unwrap();
         dims.push(v / 2);
         dims.push(2);
@@ -285,7 +293,7 @@ impl CausalSelfAttention {
         Ok(rope)
     }
 
-    fn forward(&self, x: &Tensor, freqs_cis: &Tensor) -> Result<Tensor> {
+    fn forward(&self, x: &Tensor, freqs_cis: &Tensor, block_idx: usize) -> Result<Tensor> {
         let (t, c) = x.shape().r2()?;
         let qkv = self.c_attn.forward(x)?;
         let qkv = qkv.to_dtype(DType::F32)?;
@@ -296,9 +304,31 @@ impl CausalSelfAttention {
         let target_dim = [t, self.n_head, c / self.n_head];
         let k = k.reshape(target_dim.as_slice())?.transpose(0, 1)?;
         let q = q.reshape(target_dim.as_slice())?.transpose(0, 1)?;
-        let v = v.reshape(target_dim.as_slice())?.transpose(0, 1)?;
+        let mut v = v.reshape(target_dim.as_slice())?.transpose(0, 1)?;
         let q = self.apply_rotary_emb(&q, freqs_cis)?;
-        let k = self.apply_rotary_emb(&k, freqs_cis)?;
+        let mut k = self.apply_rotary_emb(&k, freqs_cis)?;
+
+        if USE_KV_CACHE {
+            let mut cache = self.cache.kvs.lock().unwrap();
+            if let Some((cache_k, cache_v)) = &cache[block_idx] {
+                k = Tensor::cat(&[cache_k, &k], 1)?;
+                v = Tensor::cat(&[cache_v, &v], 1)?;
+                let k_seq_len = k.dims()[1];
+                if k_seq_len > CONTEXT_SIZE {
+                    k = k
+                        .narrow(1, k_seq_len - CONTEXT_SIZE, CONTEXT_SIZE)?
+                        .contiguous()?
+                }
+                let v_seq_len = v.dims()[1];
+                if v_seq_len > CONTEXT_SIZE {
+                    v = v
+                        .narrow(1, v_seq_len - CONTEXT_SIZE, CONTEXT_SIZE)?
+                        .contiguous()?
+                }
+            }
+            cache[block_idx] = Some((k.clone(), v.clone()))
+        }
+
         let k_shape = k.shape();
         let att = (q.matmul(&k.t()?)? / (*k_shape.dims().last().unwrap() as f64).sqrt())?;
         let mask = self.cache.mask(t)?.broadcast_as(att.shape())?;
@@ -330,8 +360,11 @@ impl Block {
         }
     }
 
-    fn forward(&self, x: &Tensor, freqs_cis: &Tensor) -> Result<Tensor> {
+    fn forward(&self, x: &Tensor, freqs_cis: &Tensor, block_idx: usize) -> Result<Tensor> {
-        let x = (self.attn.forward(&self.rms_1.forward(x)?, freqs_cis)? + x)?;
+        let x = (self
+            .attn
+            .forward(&self.rms_1.forward(x)?, freqs_cis, block_idx)?
+            + x)?;
         let x = (self.mlp.forward(&self.rms_2.forward(&x)?)? + x)?;
         Ok(x)
     }
@@ -358,8 +391,8 @@ impl Llama {
         // TODO: Support for mini-batches? (i.e. r2)
         let t = x.shape().r1()?;
         let mut x = self.wte.forward(x)?;
-        for block in self.blocks.iter() {
+        for (block_idx, block) in self.blocks.iter().enumerate() {
-            x = block.forward(&x, freqs_cis)?;
+            x = block.forward(&x, freqs_cis, block_idx)?;
         }
         let x = self.ln_f.forward(&x)?;
         let x = x.narrow(0, t - 1, 1)?;
@@ -400,7 +433,7 @@ struct Args {
 
     /// Use npy instead of safetensors
     #[arg(long)]
-    npy: bool,
+    npy: Option<String>,
 
     /// The temperature used to generate samples.
     #[arg(long)]
@@ -426,15 +459,16 @@ async fn main() -> Result<()> {
         Device::new_cuda(0)?
     };
     let config = Config::config_7b();
-    let cache = Cache::new(&device);
+    let cache = Cache::new(&config, &device);
     let start = std::time::Instant::now();
-    let (llama, tokenizer_filename) = if args.npy {
+    let (llama, tokenizer_filename) = match args.npy {
+        Some(npy) => {
             println!("building the model (NPY)");
-        (
+            let weights = Llama::load_npy(&device, &npy, &cache, &config)?;
-            Llama::load_npy(&device, "/data/llama.npz", &cache, &config)?,
+            let token_path = std::path::Path::new("llama-tokenizer.json").to_path_buf();
-            std::path::Path::new("llama-tokenizer.json").to_path_buf(),
+            (weights, token_path)
-        )
+        }
-    } else {
+        None => {
             let api = Api::new()?;
             let repo = Repo::new("Narsil/amall-7b".to_string(), RepoType::Model);
             println!("building the model");
@@ -453,6 +487,7 @@ async fn main() -> Result<()> {
                 Llama::load(&device, &filenames, &cache, &config)?,
                 tokenizer_filename,
             )
+        }
     };
     println!("Loaded in {:?}", start.elapsed());
     let tokenizer = Tokenizer::from_file(tokenizer_filename).map_err(E::msg)?;
@@ -470,7 +505,12 @@ async fn main() -> Result<()> {
     let start_gen = std::time::Instant::now();
     for index in 0..args.sample_len {
         let start_gen = std::time::Instant::now();
-        let ctxt = &tokens[tokens.len().saturating_sub(CONTEXT_SIZE)..];
+        let context_size = if USE_KV_CACHE && index > 0 {
+            1
+        } else {
+            CONTEXT_SIZE
+        };
+        let ctxt = &tokens[tokens.len().saturating_sub(context_size)..];
         let input = Tensor::new(ctxt, &device)?;
         let logits = llama.forward(&input, &freqs_cis)?;
 

candle-core/src/error.rs

@@ -10,6 +10,14 @@ pub enum Error {
         got: DType,
     },
 
+    #[error("invalid args for narrow: {shape:?}, dim: {dim}, start: {start}, len:{len}")]
+    NarrowInvalidArgs {
+        shape: Shape,
+        dim: usize,
+        start: usize,
+        len: usize,
+    },
+
     #[error("{op} only supports contiguous tensors")]
     RequiresContiguous { op: &'static str },
 

candle-core/src/tensor.rs

@@ -349,22 +349,35 @@ impl Tensor {
     }
 
     /// Returns a new tensor that is a narrowed version of the input, the dimension `dim`
-    /// ranges from `start` to `start + length`.
+    /// ranges from `start` to `start + len`.
-    pub fn narrow(&self, dim: usize, start: usize, length: usize) -> Result<Self> {
+    pub fn narrow(&self, dim: usize, start: usize, len: usize) -> Result<Self> {
+        let dims = self.dims();
+        if dim >= dims.len() || start + len > dims[dim] {
+            Err(Error::NarrowInvalidArgs {
+                shape: self.shape().clone(),
+                dim,
+                start,
+                len,
+            })?
+        }
+        if start == 0 && dims[dim] == len {
+            Ok(self.clone())
+        } else {
             let op = if self.track_op() {
-            Some(Op::Narrow(self.clone(), dim, start, length))
+                Some(Op::Narrow(self.clone(), dim, start, len))
             } else {
                 None
             };
             let tensor_ = Tensor_ {
                 id: TensorId::new(),
                 storage: self.storage.clone(),
-            layout: self.layout().narrow(dim, start, length)?,
+                layout: self.layout().narrow(dim, start, len)?,
                 op,
                 is_variable: false,
             };
             Ok(Tensor(Arc::new(tensor_)))
         }
+    }
 
     pub fn softmax(&self, dim: usize) -> Result<Self> {
         // TODO: unify the two branches.