Use flash-attn in gemma. (#2195)

* Use flash-attn in gemma.

* Fix flash-attn for head dim 256.

candle-examples/examples/gemma/main.rs

@@ -193,6 +193,9 @@ struct Args {
     /// The model to use.
     #[arg(long, default_value = "2b")]
     which: Which,
+
+    #[arg(long)]
+    use_flash_attn: bool,
 }
 
 fn main() -> Result<()> {
@@ -270,7 +273,7 @@ fn main() -> Result<()> {
         DType::F32
     };
     let vb = unsafe { VarBuilder::from_mmaped_safetensors(&filenames, dtype, &device)? };
-    let model = Model::new(&config, vb)?;
+    let model = Model::new(args.use_flash_attn, &config, vb)?;
 
     println!("loaded the model in {:?}", start.elapsed());
 

candle-flash-attn/kernels/flash_fwd_launch_template.h

@@ -42,6 +42,10 @@ void run_flash_fwd(Flash_fwd_params &params, cudaStream_t stream) {
                         // auto kernel = &flash_fwd_kernel<Kernel_traits, false, Is_causal, false, false, true, true, false>;
                         // printf("IsEvenMNConst = %d, IsEvenKConst = %d, Is_local = %d, Is_causal = %d, ReturnSoftmaxConst = %d, Is_dropout = %d\n", int(IsEvenMNConst), int(IsEvenKConst), int(Is_local), int(Is_causal), int(ReturnSoftmaxConst), int(Is_dropout));
                         // auto kernel = &flash_fwd_kernel<Kernel_traits, false, Is_causal, false, true, true, false>;
+                        if (smem_size >= 48 * 1024) {
+                            cudaFuncSetAttribute(
+                                kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, smem_size);
+                        }
                         // int ctas_per_sm;
                         // cudaError status_ = cudaOccupancyMaxActiveBlocksPerMultiprocessor(
                         //     &ctas_per_sm, kernel, Kernel_traits::kNThreads, smem_size);

candle-flash-attn/src/lib.rs

@@ -139,7 +139,9 @@ impl FlashAttn {
 
         let elem_count = out_shape.elem_count();
         let dst = unsafe { dev.alloc::<T>(elem_count) }.w()?;
-        let softmax_lse = dev.alloc_zeros::<f32>(b_sz * num_heads * seqlen_q).w()?;
+        let softmax_lse = dev
+            .alloc_zeros::<f32>(b_sz * 128 * num_heads * seqlen_q)
+            .w()?;
 
         let is_bf16 = if is_bf16 { 1 } else { 0 };
 

candle-transformers/src/models/gemma.rs

@@ -73,13 +73,6 @@ struct RotaryEmbedding {
     cos: Tensor,
 }
 
-fn rotate_half(xs: &Tensor) -> Result<Tensor> {
-    let last_dim = xs.dim(D::Minus1)?;
-    let xs1 = xs.narrow(D::Minus1, 0, last_dim / 2)?;
-    let xs2 = xs.narrow(D::Minus1, last_dim / 2, last_dim - last_dim / 2)?;
-    Tensor::cat(&[&xs2.neg()?, &xs1], D::Minus1)
-}
-
 impl RotaryEmbedding {
     fn new(dtype: DType, cfg: &Config, dev: &Device) -> Result<Self> {
         let dim = cfg.head_dim;
@@ -94,7 +87,6 @@ impl RotaryEmbedding {
             .to_dtype(dtype)?
             .reshape((max_seq_len, 1))?;
         let freqs = t.matmul(&inv_freq)?;
-        let freqs = Tensor::cat(&[&freqs, &freqs], D::Minus1)?;
         Ok(Self {
             sin: freqs.sin()?,
             cos: freqs.cos()?,
@@ -110,10 +102,8 @@ impl RotaryEmbedding {
         let (_b_sz, _h, seq_len, _n_embd) = q.dims4()?;
         let cos = self.cos.narrow(0, seqlen_offset, seq_len)?;
         let sin = self.sin.narrow(0, seqlen_offset, seq_len)?;
-        let cos = cos.unsqueeze(0)?.unsqueeze(0)?; // (1, 1, seq_len, dim)
-        let sin = sin.unsqueeze(0)?.unsqueeze(0)?; // (1, 1, seq_len, dim)
-        let q_embed = (q.broadcast_mul(&cos)? + rotate_half(q)?.broadcast_mul(&sin))?;
-        let k_embed = (k.broadcast_mul(&cos)? + rotate_half(k)?.broadcast_mul(&sin))?;
+        let q_embed = candle_nn::rotary_emb::rope(&q.contiguous()?, &cos, &sin)?;
+        let k_embed = candle_nn::rotary_emb::rope(&k.contiguous()?, &cos, &sin)?;
         Ok((q_embed, k_embed))
     }
 }
@@ -163,10 +153,16 @@ struct Attention {
     head_dim: usize,
     rotary_emb: Arc<RotaryEmbedding>,
     kv_cache: Option<(Tensor, Tensor)>,
+    use_flash_attn: bool,
 }
 
 impl Attention {
-    fn new(rotary_emb: Arc<RotaryEmbedding>, cfg: &Config, vb: VarBuilder) -> Result<Self> {
+    fn new(
+        rotary_emb: Arc<RotaryEmbedding>,
+        use_flash_attn: bool,
+        cfg: &Config,
+        vb: VarBuilder,
+    ) -> Result<Self> {
         let hidden_sz = cfg.hidden_size;
         let num_heads = cfg.num_attention_heads;
         let num_kv_heads = cfg.num_key_value_heads;
@@ -188,6 +184,7 @@ impl Attention {
             head_dim,
             rotary_emb,
             kv_cache: None,
+            use_flash_attn,
         })
     }
 
@@ -231,7 +228,14 @@ impl Attention {
         let value_states =
             crate::utils::repeat_kv(value_states, self.num_kv_groups)?.contiguous()?;
 
-        let attn_output = {
+        let attn_output = if self.use_flash_attn {
+            // flash-attn expects (b_sz, seq_len, nheads, head_dim)
+            let q = query_states.transpose(1, 2)?;
+            let k = key_states.transpose(1, 2)?;
+            let v = value_states.transpose(1, 2)?;
+            let scale = 1f32 / (self.head_dim as f32).sqrt();
+            flash_attn(&q, &k, &v, scale, attention_mask.is_some())?.transpose(1, 2)?
+        } else {
             let scale = 1f64 / f64::sqrt(self.head_dim as f64);
             let attn_weights = (query_states.matmul(&key_states.transpose(2, 3)?)? * scale)?;
 
@@ -253,6 +257,22 @@ impl Attention {
     }
 }
 
+#[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'")
+}
+
 #[derive(Debug, Clone)]
 struct DecoderLayer {
     self_attn: Attention,
@@ -262,8 +282,13 @@ struct DecoderLayer {
 }
 
 impl DecoderLayer {
-    fn new(rotary_emb: Arc<RotaryEmbedding>, cfg: &Config, vb: VarBuilder) -> Result<Self> {
-        let self_attn = Attention::new(rotary_emb, cfg, vb.pp("self_attn"))?;
+    fn new(
+        rotary_emb: Arc<RotaryEmbedding>,
+        use_flash_attn: bool,
+        cfg: &Config,
+        vb: VarBuilder,
+    ) -> Result<Self> {
+        let self_attn = Attention::new(rotary_emb, use_flash_attn, cfg, vb.pp("self_attn"))?;
         let mlp = MLP::new(cfg, vb.pp("mlp"))?;
         let input_layernorm =
             RmsNorm::new(cfg.hidden_size, cfg.rms_norm_eps, vb.pp("input_layernorm"))?;
@@ -312,15 +337,16 @@ pub struct Model {
 }
 
 impl Model {
-    pub fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
+    pub fn new(use_flash_attn: bool, cfg: &Config, vb: VarBuilder) -> Result<Self> {
         let vb_m = vb.pp("model");
         let embed_tokens =
             candle_nn::embedding(cfg.vocab_size, cfg.hidden_size, vb_m.pp("embed_tokens"))?;
         let rotary_emb = Arc::new(RotaryEmbedding::new(vb.dtype(), cfg, vb_m.device())?);
         let mut layers = Vec::with_capacity(cfg.num_hidden_layers);
         let vb_l = vb_m.pp("layers");
         for layer_idx in 0..cfg.num_hidden_layers {
-            let layer = DecoderLayer::new(rotary_emb.clone(), cfg, vb_l.pp(layer_idx))?;
+            let layer =
+                DecoderLayer::new(rotary_emb.clone(), use_flash_attn, cfg, vb_l.pp(layer_idx))?;
             layers.push(layer)
         }
         let norm = RmsNorm::new(cfg.hidden_size, cfg.rms_norm_eps, vb_m.pp("norm"))?;