Add a kv-cache to the quantized llama example. (#466)

* Add a kv-cache to the quantized llama example.

* Also print the prompt.

* Bugfix in q6k dequantizing.

* Another bugfix.

candle-core/src/quantized/k_quants.rs

@@ -70,7 +70,7 @@ const _: () = assert!(std::mem::size_of::<BlockQ5_1>() == 24);
 #[repr(C)]
 pub struct BlockQ8_0 {
     d: f16,
-    qs: [u8; QK8_0],
+    qs: [i8; QK8_0],
 }
 const _: () = assert!(std::mem::size_of::<BlockQ8_0>() == 34);
 
@@ -476,14 +476,14 @@ impl GgmlType for BlockQ6K {
         if k % QK_K != 0 {
             crate::bail!("dequantize_row_q6k: {k} is not divisible by {QK_K}")
         }
-        for x in xs.iter() {
+        for (idx_x, x) in xs.iter().enumerate() {
             let d = x.d.to_f32();
             let ql = &x.ql;
             let qh = &x.qh;
             let sc = &x.scales;
             for n in (0..QK_K).step_by(128) {
                 let idx = n / 128;
-                let ys = &mut ys[n..];
+                let ys = &mut ys[idx_x * QK_K + n..];
                 let sc = &sc[8 * idx..];
                 let ql = &ql[64 * idx..];
                 let qh = &qh[32 * idx..];
@@ -663,7 +663,7 @@ impl GgmlType for BlockQ8_0 {
             let id = if d != 0f32 { 1. / d } else { 0. };
             ys.d = f16::from_f32(d);
             for (y, &x) in ys.qs.iter_mut().zip(xs.iter()) {
-                *y = f32::round(x * id) as u8
+                *y = f32::round(x * id) as i8
             }
         }
         Ok(())

candle-examples/examples/ggml/main.rs

@@ -52,6 +52,7 @@ struct LayerWeights {
     head_dim: usize,
     cos: Tensor,
     sin: Tensor,
+    kv_cache: Option<(Tensor, Tensor)>,
 }
 
 fn masked_fill(on_false: &Tensor, mask: &Tensor, on_true: f32) -> Result<Tensor> {
@@ -75,7 +76,7 @@ impl LayerWeights {
         Ok(rope)
     }
 
-    fn forward_attn(&self, x: &Tensor, mask: &Tensor, index_pos: usize) -> Result<Tensor> {
+    fn forward_attn(&mut self, x: &Tensor, mask: &Tensor, index_pos: usize) -> Result<Tensor> {
         let (b_sz, seq_len, n_embd) = x.dims3()?;
         let q = self.attention_wq.forward(x)?;
         let k = self.attention_wk.forward(x)?;
@@ -94,7 +95,15 @@ impl LayerWeights {
         let q = self.apply_rotary_emb(&q, index_pos)?;
         let k = self.apply_rotary_emb(&k, index_pos)?;
 
-        // TODO: KV cache.
+        let (k, v) = match &self.kv_cache {
+            None => (k, v),
+            Some((k_cache, v_cache)) => {
+                let k = Tensor::cat(&[k_cache, &k], 2)?.contiguous()?;
+                let v = Tensor::cat(&[v_cache, &v], 2)?.contiguous()?;
+                (k, v)
+            }
+        };
+        self.kv_cache = Some((k.clone(), v.clone()));
 
         // If we start supporting MQA, we need to repeat the k and v tensors here.
 
@@ -181,6 +190,7 @@ impl ModelWeights {
                 head_dim: (ct.hparams.n_embd / ct.hparams.n_head) as usize,
                 cos: cos.clone(),
                 sin: sin.clone(),
+                kv_cache: None,
             })
         }
         Ok(Self {
@@ -209,7 +219,7 @@ impl ModelWeights {
         let (_b_sz, seq_len) = x.dims2()?;
         let mask = self.mask(seq_len)?;
         let mut layer_in = self.tok_embeddings.forward(x)?;
-        for (_layer_idx, layer) in self.layers.iter().enumerate() {
+        for layer in self.layers.iter_mut() {
             let x = layer_in;
             let residual = &x;
             let x = layer.attention_norm.forward(&x)?;
@@ -302,8 +312,9 @@ fn main() -> anyhow::Result<()> {
         .to_vec();
     let mut index_pos = 0;
     let mut logits_processor = LogitsProcessor::new(args.seed, args.temperature);
-    for _index in 0..args.sample_len {
-        let context_size = tokens.len();
+    print!("{prompt}");
+    for index in 0..args.sample_len {
+        let context_size = if index == 0 { tokens.len() } else { 1 };
         let ctxt = &tokens[tokens.len().saturating_sub(context_size)..];
         let input = Tensor::new(ctxt, &Device::Cpu)?.unsqueeze(0)?;
         let logits = model.forward(&input, index_pos)?;