Add flash-attn support. (#912)

* Add flash-attn support.

* Add the use-flash-attn flag.

* Re-enable flash-attn.

candle-transformers/src/models/wuerstchen/attention_processor.rs

@@ -11,10 +11,33 @@ pub struct Attention {
     to_out: Linear,
     heads: usize,
     scale: f64,
+    use_flash_attn: bool,
+}
+
+#[cfg(feature = "flash-attn")]
+fn flash_attn(
+    q: &Tensor,
+    k: &Tensor,
+    v: &Tensor,
+    softmax_scale: f32,
+    causal: bool,
+) -> Result<Tensor> {
+    candle_flash_attn::flash_attn(q, k, v, softmax_scale, causal)
+}
+
+#[cfg(not(feature = "flash-attn"))]
+fn flash_attn(_: &Tensor, _: &Tensor, _: &Tensor, _: f32, _: bool) -> Result<Tensor> {
+    unimplemented!("compile with '--features flash-attn'")
 }
 
 impl Attention {
-    pub fn new(query_dim: usize, heads: usize, dim_head: usize, vb: VarBuilder) -> Result<Self> {
+    pub fn new(
+        query_dim: usize,
+        heads: usize,
+        dim_head: usize,
+        use_flash_attn: bool,
+        vb: VarBuilder,
+    ) -> Result<Self> {
         let inner_dim = dim_head * heads;
         let scale = 1.0 / f64::sqrt(dim_head as f64);
         let to_q = linear(query_dim, inner_dim, vb.pp("to_q"))?;
@@ -28,6 +51,7 @@ impl Attention {
             to_out,
             scale,
             heads,
+            use_flash_attn,
         })
     }
 
@@ -62,8 +86,28 @@ impl Attention {
         let key = self.head_to_batch_dim(&key)?;
         let value = self.head_to_batch_dim(&value)?;
 
-        let attn_prs = self.get_attention_scores(&query, &key)?;
-        let xs = attn_prs.matmul(&value)?;
+        let xs = if self.use_flash_attn {
+            let init_dtype = query.dtype();
+            let q = query
+                .to_dtype(candle::DType::F16)?
+                .unsqueeze(0)?
+                .transpose(1, 2)?;
+            let k = key
+                .to_dtype(candle::DType::F16)?
+                .unsqueeze(0)?
+                .transpose(1, 2)?;
+            let v = value
+                .to_dtype(candle::DType::F16)?
+                .unsqueeze(0)?
+                .transpose(1, 2)?;
+            flash_attn(&q, &k, &v, self.scale as f32, false)?
+                .transpose(1, 2)?
+                .squeeze(0)?
+                .to_dtype(init_dtype)?
+        } else {
+            let attn_prs = self.get_attention_scores(&query, &key)?;
+            attn_prs.matmul(&value)?
+        };
         let xs = self.batch_to_head_dim(&xs)?;
 
         self.to_out

candle-transformers/src/models/wuerstchen/common.rs

@@ -174,10 +174,11 @@ impl AttnBlock {
         c_cond: usize,
         nhead: usize,
         self_attn: bool,
+        use_flash_attn: bool,
         vb: VarBuilder,
     ) -> Result<Self> {
         let norm = WLayerNorm::new(c)?;
-        let attention = Attention::new(c, nhead, c / nhead, vb.pp("attention"))?;
+        let attention = Attention::new(c, nhead, c / nhead, use_flash_attn, vb.pp("attention"))?;
         let kv_mapper_lin = candle_nn::linear(c_cond, c, vb.pp("kv_mapper.1"))?;
         Ok(Self {
             self_attn,

candle-transformers/src/models/wuerstchen/diffnext.rs

@@ -88,6 +88,7 @@ pub struct WDiffNeXt {
 }
 
 impl WDiffNeXt {
+    #[allow(clippy::too_many_arguments)]
     pub fn new(
         c_in: usize,
         c_out: usize,
@@ -95,6 +96,7 @@ impl WDiffNeXt {
         c_cond: usize,
         clip_embd: usize,
         patch_size: usize,
+        use_flash_attn: bool,
         vb: VarBuilder,
     ) -> Result<Self> {
         const C_HIDDEN: [usize; 4] = [320, 640, 1280, 1280];
@@ -169,8 +171,14 @@ impl WDiffNeXt {
                 let attn_block = if i == 0 {
                     None
                 } else {
-                    let attn_block =
-                        AttnBlock::new(c_hidden, c_cond, NHEAD[i], true, vb.pp(layer_i))?;
+                    let attn_block = AttnBlock::new(
+                        c_hidden,
+                        c_cond,
+                        NHEAD[i],
+                        true,
+                        use_flash_attn,
+                        vb.pp(layer_i),
+                    )?;
                     layer_i += 1;
                     Some(attn_block)
                 };
@@ -208,8 +216,14 @@ impl WDiffNeXt {
                 let attn_block = if i == 0 {
                     None
                 } else {
-                    let attn_block =
-                        AttnBlock::new(c_hidden, c_cond, NHEAD[i], true, vb.pp(layer_i))?;
+                    let attn_block = AttnBlock::new(
+                        c_hidden,
+                        c_cond,
+                        NHEAD[i],
+                        true,
+                        use_flash_attn,
+                        vb.pp(layer_i),
+                    )?;
                     layer_i += 1;
                     Some(attn_block)
                 };

candle-transformers/src/models/wuerstchen/prior.rs

@@ -21,6 +21,7 @@ pub struct WPrior {
 }
 
 impl WPrior {
+    #[allow(clippy::too_many_arguments)]
     pub fn new(
         c_in: usize,
         c: usize,
@@ -28,6 +29,7 @@ impl WPrior {
         c_r: usize,
         depth: usize,
         nhead: usize,
+        use_flash_attn: bool,
         vb: VarBuilder,
     ) -> Result<Self> {
         let projection = candle_nn::conv2d(c_in, c, 1, Default::default(), vb.pp("projection"))?;
@@ -44,6 +46,7 @@ impl WPrior {
                 c,
                 nhead,
                 true,
+                use_flash_attn,
                 vb.pp(format!("blocks.{}", 3 * index + 2)),
             )?;
             blocks.push(Block {