GQA support in the quantized model. (#555)

* GQA support in the quantized model.

* Fix the reshaping.

* Fix the main llama model.

* Infer the proper gqa from the model kind.

candle-examples/examples/llama/model.rs

@@ -291,7 +291,7 @@ impl CausalSelfAttention {
             let x = x
                 .unsqueeze(2)?
                 .expand((b_sz, n_kv_head, n_rep, seq_len, head_dim))?
-                .reshape((b_sz, n_kv_head, n_rep, seq_len, head_dim))?;
+                .reshape((b_sz, n_kv_head * n_rep, seq_len, head_dim))?;
             Ok(x)
         }
     }

candle-examples/examples/quantized/main.rs

@@ -68,6 +68,7 @@ struct LayerWeights {
     feed_forward_w3: QMatMul,
     ffn_norm: RmsNorm,
     n_head: usize,
+    n_kv_head: usize,
     head_dim: usize,
     cos: Tensor,
     sin: Tensor,
@@ -125,10 +126,10 @@ impl LayerWeights {
             .reshape((b_sz, seq_len, self.n_head, self.head_dim))?
             .transpose(1, 2)?;
         let k = k
-            .reshape((b_sz, seq_len, self.n_head, self.head_dim))?
+            .reshape((b_sz, seq_len, self.n_kv_head, self.head_dim))?
             .transpose(1, 2)?;
         let v = v
-            .reshape((b_sz, seq_len, self.n_head, self.head_dim))?
+            .reshape((b_sz, seq_len, self.n_kv_head, self.head_dim))?
             .transpose(1, 2)?;
 
         let q = self.apply_rotary_emb(&q, index_pos)?;
@@ -144,7 +145,9 @@ impl LayerWeights {
         };
         self.kv_cache = Some((k.clone(), v.clone()));
 
-        // If we start supporting MQA, we need to repeat the k and v tensors here.
+        // Support for MQA, useful for 70B models.
+        let k = self.repeat_kv(k)?;
+        let v = self.repeat_kv(v)?;
 
         let att = (q.matmul(&k.t()?)? / (self.head_dim as f64).sqrt())?;
         let mask = mask.broadcast_as(att.shape())?;
@@ -156,6 +159,20 @@ impl LayerWeights {
         let y = self.attention_wo.forward(&y)?;
         Ok(y)
     }
+
+    fn repeat_kv(&self, x: Tensor) -> Result<Tensor> {
+        let n_rep = self.n_head / self.n_kv_head;
+        if n_rep == 1 {
+            Ok(x)
+        } else {
+            let (b_sz, n_kv_head, seq_len, head_dim) = x.dims4()?;
+            let x = x
+                .unsqueeze(2)?
+                .expand((b_sz, n_kv_head, n_rep, seq_len, head_dim))?
+                .reshape((b_sz, n_kv_head * n_rep, seq_len, head_dim))?;
+            Ok(x)
+        }
+    }
 }
 
 struct ModelWeights {
@@ -179,7 +196,7 @@ impl WeightMap {
 }
 
 impl ModelWeights {
-    fn new(mut ct: Content) -> Result<Self> {
+    fn new(mut ct: Content, gqa: usize) -> Result<Self> {
         let cpu = &Device::Cpu;
         let head_dim = (ct.hparams.n_embd / ct.hparams.n_head) as usize;
 
@@ -226,6 +243,7 @@ impl ModelWeights {
                 feed_forward_w3: QMatMul::from_qtensor(feed_forward_w3),
                 ffn_norm: RmsNorm::new(ffn_norm)?,
                 n_head: ct.hparams.n_head as usize,
+                n_kv_head: ct.hparams.n_head as usize / gqa,
                 head_dim: (ct.hparams.n_embd / ct.hparams.n_head) as usize,
                 cos: cos.clone(),
                 sin: sin.clone(),
@@ -347,6 +365,10 @@ struct Args {
     /// The model size to use.
     #[arg(long, default_value = "7b")]
     which: Which,
+
+    /// Group-Query Attention, use 8 for the 70B version of LLaMAv2.
+    #[arg(long)]
+    gqa: Option<usize>,
 }
 
 impl Args {
@@ -468,7 +490,11 @@ fn main() -> anyhow::Result<()> {
         start.elapsed().as_secs_f32(),
     );
     println!("params: {:?}", model.hparams);
-    let mut model = ModelWeights::new(model)?;
+    let default_gqa = match args.which {
+        Which::L7b | Which::L13b => 1,
+        Which::L70b => 8,
+    };
+    let mut model = ModelWeights::new(model, args.gqa.unwrap_or(default_gqa))?;
     println!("model built");
 
     let tokenizer = args.tokenizer()?;