Lining up the flash attn version with the non-flash one. (#248)

* Move the flash-attn function in the proper crate.

* Causality tweak.

candle-examples/examples/llama/model.rs

@@ -146,19 +146,18 @@ struct CausalSelfAttention {
 }
 
 #[cfg(feature = "flash-attn")]
-fn flash_attn(softmax_scale: f32, q: &Tensor, k: &Tensor, v: &Tensor) -> Result<Tensor> {
+fn flash_attn(
-    q.custom_op3(
+    q: &Tensor,
-        k,
+    k: &Tensor,
-        v,
+    v: &Tensor,
-        candle_flash_attn::FlashHdim32Sm80 {
+    softmax_scale: f32,
-            softmax_scale,
+    causal: bool,
-            causal: true,
+) -> Result<Tensor> {
-        },
+    candle_flash_attn::flash_attn(q, k, v, softmax_scale, causal)
-    )
 }
 
 #[cfg(not(feature = "flash-attn"))]
-fn flash_attn(_: f32, _: &Tensor, _: &Tensor, _: &Tensor) -> Result<Tensor> {
+fn flash_attn(_: &Tensor, _: &Tensor, _: &Tensor, _: f32, _: bool) -> Result<Tensor> {
     unimplemented!("compile with '--features flash-attn'")
 }
 
@@ -225,7 +224,7 @@ impl CausalSelfAttention {
             let k = k.transpose(1, 2)?;
             let v = v.transpose(1, 2)?;
             let softmax_scale = 1f32 / (self.head_dim as f32).sqrt();
-            flash_attn(softmax_scale, &q, &k, &v)?.transpose(1, 2)?
+            flash_attn(&q, &k, &v, softmax_scale, seq_len > 1)?.transpose(1, 2)?
         } else {
             let in_dtype = q.dtype();
             let q = q.to_dtype(DType::F32)?;

candle-flash-attn/src/lib.rs

@@ -3,7 +3,7 @@ mod ffi;
 use candle::backend::BackendStorage;
 use candle::cuda_backend::cudarc::driver::DevicePtr;
 use candle::cuda_backend::WrapErr;
-use candle::{CpuStorage, Error, Layout, Result, Shape};
+use candle::{CpuStorage, Error, Layout, Result, Shape, Tensor};
 use half::f16;
 
 pub struct FlashHdim32Sm80 {
@@ -144,3 +144,20 @@ impl candle::CustomOp3 for FlashHdim32Sm80 {
         Ok((dst, out_shape))
     }
 }
+
+pub fn flash_attn(
+    q: &Tensor,
+    k: &Tensor,
+    v: &Tensor,
+    softmax_scale: f32,
+    causal: bool,
+) -> Result<Tensor> {
+    q.custom_op3(
+        k,
+        v,
+        FlashHdim32Sm80 {
+            softmax_scale,
+            causal,
+        },
+    )
+}