More rotating kv-cache.

candle-transformers/src/models/mimi/transformer.rs

@@ -216,16 +216,6 @@ impl StreamingMultiheadAttention {
  let pre_ws = match mask {
  None => pre_ws,
  Some(mask) => {
- // This is a bit cumbersome and slightly incorrect: when providing a new slice
- // the kv cache will have a slice offset rather than offset + t. In the mimi
- // context of an offset of 250, this would not make much difference though.
- let mask_len = mask.dim(D::Minus1)?;
- let pre_ws_len = pre_ws.dim(D::Minus1)?;
- let mask = if pre_ws_len < mask_len {
- mask.narrow(D::Minus1, mask_len - pre_ws_len, pre_ws_len)?
- } else {
- mask.clone()
- };
  let mask = mask.broadcast_left((b, self.num_heads))?;
  let neg_inf = self.neg_inf.broadcast_as(pre_ws.shape())?;
  mask.where_cond(&neg_inf, &pre_ws)?
@@ -639,18 +629,22 @@ impl StreamingTransformer {
 
  pub fn forward_ca(&mut self, xs: &Tensor, ca_src: Option<&Tensor>) -> Result<Tensor> {
  let (_b, t, c) = xs.dims3()?;
- // We will extract at most "context" from the kv_cache.
- // Note that the mask will discard the values that are before context.
  let pos = self.layers[0]
  .self_attn
  .kv_cache
  .k_cache()
- .current_seq_len()
- .min(self.context);
+ .current_seq_len();
  let mask = if t == 1 {
  None
  } else {
- Some(get_mask(t, pos + t, self.context, xs.device())?)
+ let cache_out_len = if t < self.context {
+ (pos + t).min(self.context)
+ } else {
+ t
+ };
+ // TODO: this is wrong, the mask depends on the kv-cache offset because of its rotating
+ // nature.
+ Some(get_mask(t, cache_out_len, self.context, xs.device())?)
  };
  let mut xs = match self.positional_embedding {
  PositionalEmbedding::Rope | PositionalEmbedding::None => xs.clone(),