More segment-anything again. (#764)

* More segment-anything again.

* Transformer block forward.

* Two-ways transformer.

* Position embeddings.

* Sketch the prompt encoder.

* More prompt-encoder.

* More prompt-encoder.

* Add the main sam module.

* Embed the transformer.

* And hook the transformer forward step.

* Build the model.

* Handle the global attn indexes.

* Get the model to load.

candle-examples/examples/segment-anything/main.rs

@@ -8,9 +8,11 @@ extern crate intel_mkl_src;
 #[cfg(feature = "accelerate")]
 extern crate accelerate_src;
 
-mod model_image_encoder;
-mod model_mask_decoder;
-mod model_transformer;
+pub mod model_image_encoder;
+pub mod model_mask_decoder;
+pub mod model_prompt_encoder;
+pub mod model_sam;
+pub mod model_transformer;
 
 use candle::{DType, IndexOp, Result, Tensor, D};
 use candle_nn::{layer_norm, LayerNorm, Linear, Module, VarBuilder};
@@ -82,7 +84,7 @@ impl Module for MlpBlock {
 #[derive(Parser)]
 struct Args {
     #[arg(long)]
-    model: Option<String>,
+    model: String,
 
     #[arg(long)]
     image: String,
@@ -95,10 +97,15 @@ struct Args {
 pub fn main() -> anyhow::Result<()> {
     let args = Args::parse();
 
-    let _device = candle_examples::device(args.cpu)?;
+    let device = candle_examples::device(args.cpu)?;
 
-    let image = candle_examples::imagenet::load_image224(args.image)?;
+    let image = candle_examples::imagenet::load_image224(args.image)?.to_device(&device);
     println!("loaded image {image:?}");
 
+    let weights = unsafe { candle::safetensors::MmapedFile::new(args.model)? };
+    let weights = weights.deserialize()?;
+    let vb = VarBuilder::from_safetensors(vec![weights], DType::F32, &device);
+    let _sam = model_sam::Sam::new(768, 12, 12, &[2, 5, 8, 11], vb)?; // sam_vit_b
+
     Ok(())
 }

candle-examples/examples/segment-anything/model_image_encoder.rs

@@ -47,7 +47,7 @@ impl Attention {
         num_heads: usize,
         qkv_bias: bool,
         use_rel_pos: bool,
-        window_size: usize,
+        input_size: (usize, usize),
         vb: VarBuilder,
     ) -> Result<Self> {
         let qkv = crate::linear(vb.pp("qkv"), dim, dim * 3, qkv_bias)?;
@@ -55,8 +55,8 @@ impl Attention {
         let head_dim = dim / num_heads;
         let scale = 1. / (head_dim as f64).sqrt();
         let rel_pos_hw = if use_rel_pos {
-            let h = vb.get((2 * window_size - 1, head_dim), "rel_pos_h")?;
-            let w = vb.get((2 * window_size - 1, head_dim), "rel_pos_w")?;
+            let h = vb.get((2 * input_size.0 - 1, head_dim), "rel_pos_h")?;
+            let w = vb.get((2 * input_size.1 - 1, head_dim), "rel_pos_w")?;
             Some((h, w))
         } else {
             None
@@ -114,16 +114,22 @@ impl Block {
         qkv_bias: bool,
         use_rel_pos: bool,
         window_size: usize,
+        input_size: (usize, usize),
         vb: VarBuilder,
     ) -> Result<Self> {
         let norm1 = layer_norm(dim, 1e-5, vb.pp("norm1"))?;
         let norm2 = layer_norm(dim, 1e-5, vb.pp("norm2"))?;
+        let input_size_attn = if window_size == 0 {
+            input_size
+        } else {
+            (window_size, window_size)
+        };
         let attn = Attention::new(
             dim,
             num_heads,
             qkv_bias,
             use_rel_pos,
-            window_size,
+            input_size_attn,
             vb.pp("attn"),
         )?;
         let mlp = crate::MlpBlock::new(dim, dim * 4, vb.pp("mlp"))?;
@@ -154,7 +160,7 @@ impl Module for Block {
 }
 
 #[derive(Debug)]
-struct ImageEncoderViT {
+pub struct ImageEncoderViT {
     img_size: usize,
     patch_embed: PatchEmbed,
     blocks: Vec<Block>,
@@ -167,7 +173,7 @@ struct ImageEncoderViT {
 
 impl ImageEncoderViT {
     #[allow(clippy::too_many_arguments)]
-    fn new(
+    pub fn new(
         img_size: usize,
         patch_size: usize,
         in_chans: usize,
@@ -179,6 +185,7 @@ impl ImageEncoderViT {
         use_rel_pos: bool,
         use_abs_pos: bool,
         window_size: usize,
+        global_attn_indexes: &[usize],
         vb: VarBuilder,
     ) -> Result<Self> {
         let patch_embed = PatchEmbed::new(
@@ -192,12 +199,18 @@ impl ImageEncoderViT {
         let mut blocks = Vec::with_capacity(depth);
         let vb_b = vb.pp("blocks");
         for i in 0..depth {
+            let window_size = if global_attn_indexes.contains(&i) {
+                0
+            } else {
+                window_size
+            };
             let block = Block::new(
                 embed_dim,
                 num_heads,
                 qkv_bias,
                 use_rel_pos,
                 window_size,
+                (img_size / patch_size, img_size / patch_size),
                 vb_b.pp(i),
             )?;
             blocks.push(block)

candle-examples/examples/segment-anything/model_mask_decoder.rs

@@ -1,6 +1,8 @@
 use candle::{DType, IndexOp, Result, Tensor, D};
 use candle_nn::{layer_norm, LayerNorm, Linear, Module, VarBuilder};
 
+use crate::model_transformer::TwoWayTransformer;
+
 #[derive(Debug)]
 struct MlpMaskDecoder {
     layers: Vec<Linear>,
@@ -53,7 +55,7 @@ impl Module for MlpMaskDecoder {
 }
 
 #[derive(Debug)]
-struct MaskDecoder {
+pub struct MaskDecoder {
     iou_token: candle_nn::Embedding,
     mask_tokens: candle_nn::Embedding,
     iou_prediction_head: MlpMaskDecoder,
@@ -62,17 +64,18 @@ struct MaskDecoder {
     output_upscaling_conv2: candle_nn::ConvTranspose2d,
     num_mask_tokens: usize,
     output_hypernetworks_mlps: Vec<MlpMaskDecoder>,
+    transformer: TwoWayTransformer,
 }
 
 impl MaskDecoder {
-    fn new(
+    pub fn new(
         transformer_dim: usize,
         num_multimask_outputs: usize,
         iou_head_depth: usize,
         iou_head_hidden_dim: usize,
         vb: VarBuilder,
     ) -> Result<Self> {
-        let num_mask_tokens = num_multimask_outputs - 1;
+        let num_mask_tokens = num_multimask_outputs + 1;
         let iou_prediction_head = MlpMaskDecoder::new(
             transformer_dim,
             iou_head_hidden_dim,
@@ -117,6 +120,13 @@ impl MaskDecoder {
             )?;
             output_hypernetworks_mlps.push(mlp)
         }
+        let transformer = TwoWayTransformer::new(
+            /* depth */ 2,
+            /* embedding_dim */ transformer_dim,
+            /* num_heads */ 8,
+            /* mlp_dim */ 2048,
+            vb.pp("transformer"),
+        )?;
         Ok(Self {
             iou_token,
             mask_tokens,
@@ -126,6 +136,7 @@ impl MaskDecoder {
             output_upscaling_conv2,
             num_mask_tokens,
             output_hypernetworks_mlps,
+            transformer,
         })
     }
 
@@ -182,7 +193,7 @@ impl MaskDecoder {
         let (b, c, h, w) = src.dims4()?;
 
         // Run the transformer
-        let (hs, src) = run_transformer(&src, &pos_src, &tokens)?;
+        let (hs, src) = self.transformer.forward(&src, &pos_src, &tokens)?;
         let iou_token_out = hs.i((.., 0))?;
         let mask_tokens_out = hs.i((.., 1, 1 + self.num_mask_tokens))?;
 
@@ -216,7 +227,3 @@ impl MaskDecoder {
 fn repeat_interleave(_img: &Tensor, _repeats: usize, _dim: usize) -> Result<Tensor> {
     todo!()
 }
-
-fn run_transformer(_src: &Tensor, _pos: &Tensor, _tokens: &Tensor) -> Result<(Tensor, Tensor)> {
-    todo!()
-}

candle-examples/examples/segment-anything/model_prompt_encoder.rs

@@ -0,0 +1,192 @@
+use candle::{DType, IndexOp, Result, Tensor, D};
+use candle_nn::{layer_norm, LayerNorm, Linear, Module, VarBuilder};
+
+#[derive(Debug)]
+struct PostionEmbeddingRandom {
+    positional_encoding_gaussian_matrix: Tensor,
+}
+
+impl PostionEmbeddingRandom {
+    fn new(num_pos_feats: usize, vb: VarBuilder) -> Result<Self> {
+        let positional_encoding_gaussian_matrix =
+            vb.get((2, num_pos_feats), "positional_encoding_gaussian_matrix")?;
+        Ok(Self {
+            positional_encoding_gaussian_matrix,
+        })
+    }
+
+    fn pe_encoding(&self, coords: &Tensor) -> Result<Tensor> {
+        let coords = coords.affine(2., -1.)?;
+        let coords = coords.matmul(&self.positional_encoding_gaussian_matrix)?;
+        let coords = (coords * (2. * std::f64::consts::PI))?;
+        Tensor::cat(&[coords.sin()?, coords.cos()?], D::Minus1)
+    }
+
+    fn forward(&self, h: usize, w: usize) -> Result<Tensor> {
+        let device = self.positional_encoding_gaussian_matrix.device();
+        let grid = Tensor::ones((h, w), DType::F32, device)?;
+        // TODO: cumsum
+        let x_embed = (&grid - 0.5)?;
+        // TODO: cumsum
+        let y_embed = (&grid - 0.5)?;
+        let x_embed = (x_embed / w as f64)?;
+        let y_embed = (y_embed / h as f64)?;
+        let coords = Tensor::stack(&[&x_embed, &y_embed], D::Minus1)?;
+        self.pe_encoding(&coords)?.permute((2, 0, 1))
+    }
+
+    fn forward_with_coords(
+        &self,
+        coords_input: &Tensor,
+        image_size: (usize, usize),
+    ) -> Result<Tensor> {
+        let coords0 = (coords_input.narrow(D::Minus1, 0, 1)? / image_size.1 as f64)?;
+        let coords1 = (coords_input.narrow(D::Minus1, 1, 1)? / image_size.0 as f64)?;
+        let c = coords_input.dim(D::Minus1)?;
+        let coords_rest = coords_input.narrow(D::Minus1, 2, c - 2)?;
+        let coords = Tensor::cat(&[&coords0, &coords1, &coords_rest], D::Minus1)?;
+        self.pe_encoding(&coords)
+    }
+}
+
+#[derive(Debug)]
+pub struct PromptEncoder {
+    pe_layer: PostionEmbeddingRandom,
+    point_embeddings: Vec<candle_nn::Embedding>,
+    not_a_point_embed: candle_nn::Embedding,
+    mask_downscaling_conv1: candle_nn::Conv2d,
+    mask_downscaling_ln1: LayerNorm,
+    mask_downscaling_conv2: candle_nn::Conv2d,
+    mask_downscaling_ln2: LayerNorm,
+    mask_downscaling_conv3: candle_nn::Conv2d,
+    no_mask_embed: candle_nn::Embedding,
+    image_embedding_size: (usize, usize),
+    input_image_size: (usize, usize),
+}
+
+impl PromptEncoder {
+    pub fn new(
+        embed_dim: usize,
+        image_embedding_size: (usize, usize),
+        input_image_size: (usize, usize),
+        mask_in_chans: usize,
+        vb: VarBuilder,
+    ) -> Result<Self> {
+        let num_points_embeddings = 4;
+        let pe_layer = PostionEmbeddingRandom::new(embed_dim / 2, vb.pp("pe_layer"))?;
+        let not_a_point_embed = candle_nn::embedding(1, embed_dim, vb.pp("not_a_point_embed"))?;
+        let no_mask_embed = candle_nn::embedding(1, embed_dim, vb.pp("no_mask_embed"))?;
+        let cfg = candle_nn::Conv2dConfig {
+            stride: 2,
+            ..Default::default()
+        };
+        let mask_downscaling_conv1 =
+            candle_nn::conv2d(1, mask_in_chans / 4, 2, cfg, vb.pp("mask_downscaling.0"))?;
+        let mask_downscaling_conv2 = candle_nn::conv2d(
+            mask_in_chans / 4,
+            mask_in_chans,
+            2,
+            cfg,
+            vb.pp("mask_downscaling.3"),
+        )?;
+        let mask_downscaling_conv3 = candle_nn::conv2d(
+            mask_in_chans,
+            embed_dim,
+            1,
+            Default::default(),
+            vb.pp("mask_downscaling.6"),
+        )?;
+        let mask_downscaling_ln1 =
+            layer_norm(mask_in_chans / 4, 1e-6, vb.pp("mask_downscaling.1"))?;
+        let mask_downscaling_ln2 = layer_norm(mask_in_chans, 1e-6, vb.pp("mask_downscaling.4"))?;
+        let mut point_embeddings = Vec::with_capacity(num_points_embeddings);
+        let vb_e = vb.pp("point_embeddings");
+        for i in 0..num_points_embeddings {
+            let emb = candle_nn::embedding(1, embed_dim, vb_e.pp(i))?;
+            point_embeddings.push(emb)
+        }
+        Ok(Self {
+            pe_layer,
+            point_embeddings,
+            not_a_point_embed,
+            mask_downscaling_conv1,
+            mask_downscaling_ln1,
+            mask_downscaling_conv2,
+            mask_downscaling_ln2,
+            mask_downscaling_conv3,
+            no_mask_embed,
+            image_embedding_size,
+            input_image_size,
+        })
+    }
+
+    fn embed_masks(&self, masks: &Tensor) -> Result<Tensor> {
+        masks
+            .apply(&self.mask_downscaling_conv1)?
+            .apply(&self.mask_downscaling_ln1)?
+            .gelu()?
+            .apply(&self.mask_downscaling_conv2)?
+            .apply(&self.mask_downscaling_ln2)?
+            .gelu()?
+            .apply(&self.mask_downscaling_conv3)
+    }
+
+    fn embed_points(&self, points: &Tensor, labels: &Tensor, pad: bool) -> Result<Tensor> {
+        let points = (points + 0.5)?;
+        let points = if pad { todo!() } else { points };
+        let point_embedding = self
+            .pe_layer
+            .forward_with_coords(&points, self.input_image_size)?;
+        // TODO: tweak based on labels.
+        Ok(point_embedding)
+    }
+
+    fn embed_boxes(&self, boxes: &Tensor) -> Result<Tensor> {
+        let boxes = (boxes + 0.5)?;
+        let coords = boxes.reshape((boxes.elem_count() / 4, 2, 2))?;
+        let corner_embedding = self
+            .pe_layer
+            .forward_with_coords(&coords, self.input_image_size)?;
+        let ce1 = corner_embedding.i((.., 0))?;
+        let ce2 = corner_embedding.i((.., 1))?;
+        let ce1 = (ce1 + self.point_embeddings[2].embeddings())?;
+        let ce2 = (ce2 + self.point_embeddings[3].embeddings())?;
+        Tensor::cat(&[&ce1, &ce2], 1)
+    }
+
+    fn forward(
+        &self,
+        points: Option<(&Tensor, &Tensor)>,
+        boxes: Option<&Tensor>,
+        masks: Option<&Tensor>,
+    ) -> Result<(Tensor, Tensor)> {
+        let se_points = match points {
+            Some((coords, labels)) => Some(self.embed_points(coords, labels, boxes.is_none())?),
+            None => None,
+        };
+        let se_boxes = match boxes {
+            Some(boxes) => Some(self.embed_boxes(boxes)?),
+            None => None,
+        };
+        let sparse_embeddings = match (se_points, se_boxes) {
+            (Some(se_points), Some(se_boxes)) => Tensor::cat(&[se_points, se_boxes], 1)?,
+            (Some(se_points), None) => se_points,
+            (None, Some(se_boxes)) => se_boxes,
+            (None, None) => Tensor::zeros(1, DType::F32, &candle::Device::Cpu)?,
+        };
+
+        let dense_embeddings = match masks {
+            None => {
+                let emb = self.no_mask_embed.embeddings();
+                emb.reshape((1, emb.elem_count(), 1, 1))?.expand((
+                    1,
+                    0,
+                    self.image_embedding_size.0,
+                    self.image_embedding_size.1,
+                ))?
+            }
+            Some(masks) => self.embed_masks(masks)?,
+        };
+        Ok((sparse_embeddings, dense_embeddings))
+    }
+}

candle-examples/examples/segment-anything/model_sam.rs

@@ -0,0 +1,72 @@
+use candle::{DType, IndexOp, Result, Tensor, D};
+use candle_nn::{layer_norm, LayerNorm, Linear, Module, VarBuilder};
+
+use crate::model_image_encoder::ImageEncoderViT;
+use crate::model_mask_decoder::MaskDecoder;
+use crate::model_prompt_encoder::PromptEncoder;
+
+#[derive(Debug)]
+pub struct Sam {
+    image_encoder: ImageEncoderViT,
+    prompt_encoder: PromptEncoder,
+    mask_decoder: MaskDecoder,
+    pixel_mean: Tensor,
+    pixel_std: Tensor,
+}
+
+impl Sam {
+    pub fn new(
+        encoder_embed_dim: usize,
+        encoder_depth: usize,
+        encoder_num_heads: usize,
+        encoder_global_attn_indexes: &[usize],
+        vb: VarBuilder,
+    ) -> Result<Self> {
+        const PROMPT_EMBED_DIM: usize = 256;
+        const IMAGE_SIZE: usize = 1024;
+        const VIT_PATCH_SIZE: usize = 16;
+
+        let image_embedding_size = IMAGE_SIZE / VIT_PATCH_SIZE;
+
+        let image_encoder = ImageEncoderViT::new(
+            IMAGE_SIZE,
+            VIT_PATCH_SIZE,
+            3,
+            encoder_embed_dim,
+            encoder_depth,
+            encoder_num_heads,
+            PROMPT_EMBED_DIM,
+            /* qkv_bias */ true,
+            /* use_rel_pos */ true,
+            /* use_abs_pos */ true,
+            /* window_size */ 14,
+            /* global_attn_indexes */ encoder_global_attn_indexes,
+            vb.pp("image_encoder"),
+        )?;
+        let prompt_encoder = PromptEncoder::new(
+            PROMPT_EMBED_DIM,
+            (image_embedding_size, image_embedding_size),
+            (IMAGE_SIZE, IMAGE_SIZE),
+            16,
+            vb.pp("prompt_encoder"),
+        )?;
+        let mask_decoder = MaskDecoder::new(
+            PROMPT_EMBED_DIM,
+            /* num_multitask_outputs */ 3,
+            /* iou_head_depth */ 3,
+            /* iou_head_hidden_dim */ 256,
+            vb.pp("mask_decoder"),
+        )?;
+        let pixel_mean =
+            Tensor::new(&[123.675f32, 116.28, 103.53], vb.device())?.reshape((3, 1, 1))?;
+        let pixel_std =
+            Tensor::new(&[58.395f32, 57.12, 57.375], vb.device())?.reshape((3, 1, 1))?;
+        Ok(Self {
+            image_encoder,
+            prompt_encoder,
+            mask_decoder,
+            pixel_std,
+            pixel_mean,
+        })
+    }
+}

candle-examples/examples/segment-anything/model_transformer.rs

@@ -75,3 +75,146 @@ struct TwoWayAttentionBlock {
     cross_attn_image_to_token: Attention,
     skip_first_layer_pe: bool,
 }
+
+impl TwoWayAttentionBlock {
+    fn new(
+        embedding_dim: usize,
+        num_heads: usize,
+        mlp_dim: usize,
+        skip_first_layer_pe: bool,
+        vb: VarBuilder,
+    ) -> Result<Self> {
+        let self_attn = Attention::new(embedding_dim, num_heads, 1, vb.pp("self_attn"))?;
+        let norm1 = layer_norm(embedding_dim, 1e-5, vb.pp("norm1"))?;
+        let norm2 = layer_norm(embedding_dim, 1e-5, vb.pp("norm2"))?;
+        let norm3 = layer_norm(embedding_dim, 1e-5, vb.pp("norm3"))?;
+        let norm4 = layer_norm(embedding_dim, 1e-5, vb.pp("norm4"))?;
+        let self_attn = Attention::new(embedding_dim, num_heads, 1, vb.pp("self_attn"))?;
+        let cross_attn_token_to_image = Attention::new(
+            embedding_dim,
+            num_heads,
+            2,
+            vb.pp("cross_attn_token_to_image"),
+        )?;
+        let cross_attn_image_to_token = Attention::new(
+            embedding_dim,
+            num_heads,
+            2,
+            vb.pp("cross_attn_image_to_token"),
+        )?;
+        // TODO: use relu in this mlp
+        let mlp = crate::MlpBlock::new(embedding_dim, mlp_dim, vb.pp("mlp"))?;
+        Ok(Self {
+            self_attn,
+            norm1,
+            cross_attn_image_to_token,
+            norm2,
+            mlp,
+            norm3,
+            norm4,
+            cross_attn_token_to_image,
+            skip_first_layer_pe,
+        })
+    }
+
+    fn forward(
+        &self,
+        queries: &Tensor,
+        keys: &Tensor,
+        query_pe: &Tensor,
+        key_pe: &Tensor,
+    ) -> Result<(Tensor, Tensor)> {
+        // Self attention block
+        let queries = if self.skip_first_layer_pe {
+            self.self_attn.forward(queries, keys, queries)?
+        } else {
+            let q = (queries + query_pe)?;
+            let attn_out = self.self_attn.forward(&q, &q, queries)?;
+            (queries + attn_out)?
+        };
+        let queries = self.norm1.forward(&queries)?;
+
+        // Cross attention block, tokens attending to image embedding
+        let q = (&queries + query_pe)?;
+        let k = (keys + key_pe)?;
+        let attn_out = self.cross_attn_token_to_image.forward(&q, &k, keys)?;
+        let queries = (&queries + attn_out)?;
+        let queries = self.norm2.forward(&queries)?;
+
+        // MLP block
+        let mlp_out = self.mlp.forward(&queries);
+        let queries = (queries + mlp_out)?;
+        let queries = self.norm3.forward(&queries)?;
+
+        // Cross attention block, image embedding attending to tokens
+        let q = (&queries + query_pe)?;
+        let k = (keys + key_pe)?;
+        let attn_out = self.cross_attn_image_to_token.forward(&k, &q, &queries)?;
+        let keys = (keys + attn_out)?;
+        let keys = self.norm4.forward(&keys)?;
+
+        Ok((queries, keys))
+    }
+}
+
+#[derive(Debug)]
+pub struct TwoWayTransformer {
+    layers: Vec<TwoWayAttentionBlock>,
+    final_attn_token_to_image: Attention,
+    norm_final_attn: LayerNorm,
+}
+
+impl TwoWayTransformer {
+    pub fn new(
+        depth: usize,
+        embedding_dim: usize,
+        num_heads: usize,
+        mlp_dim: usize,
+        vb: VarBuilder,
+    ) -> Result<Self> {
+        let vb_l = vb.pp("layers");
+        let mut layers = Vec::with_capacity(depth);
+        for i in 0..depth {
+            let layer =
+                TwoWayAttentionBlock::new(embedding_dim, num_heads, mlp_dim, i == 0, vb_l.pp(i))?;
+            layers.push(layer)
+        }
+        let final_attn_token_to_image = Attention::new(
+            embedding_dim,
+            num_heads,
+            2,
+            vb.pp("final_attn_token_to_image"),
+        )?;
+        let norm_final_attn = layer_norm(embedding_dim, 1e-5, vb.pp("norm_final_attn"))?;
+        Ok(Self {
+            layers,
+            final_attn_token_to_image,
+            norm_final_attn,
+        })
+    }
+
+    pub fn forward(
+        &self,
+        image_embedding: &Tensor,
+        image_pe: &Tensor,
+        point_embedding: &Tensor,
+    ) -> Result<(Tensor, Tensor)> {
+        let (bs, c, h, w) = image_embedding.dims4()?;
+        let image_embedding = image_embedding.flatten_from(2)?.permute((0, 2, 1))?;
+        let image_pe = image_pe.flatten_from(2)?.permute((0, 2, 1))?;
+
+        let mut queries = point_embedding.clone();
+        let mut keys = image_embedding;
+
+        for layer in self.layers.iter() {
+            (queries, keys) = layer.forward(&queries, &keys, point_embedding, &image_pe)?
+        }
+
+        let q = (&queries + point_embedding)?;
+        let k = (&keys + image_pe)?;
+        let attn_out = self.final_attn_token_to_image.forward(&q, &k, &keys)?;
+        let queries = (queries + attn_out)?.apply(&self.norm_final_attn)?;
+
+        Ok((queries, keys))
+    }
+}