phase 5: naive multi-head attention

- Batched GEMM via cublasGemmStridedBatchedEx
- Causal mask CUDA kernel (F32 + BF16)
- Element-wise scale CUDA kernel (F32 + BF16)
- attention() composing: batched_matmul + scale + causal_mask + softmax
- Fixed to_device/contiguous infinite recursion (GPU contiguous via CPU round-trip)
- 5 attention tests passing (max_err < 3e-7 F32)
- Total: 61 tests passing across all crates

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

crates/xserv-kernels/src/attention.rs

@@ -0,0 +1,77 @@
+use std::ffi::c_void;
+use xserv_tensor::{DType, Tensor};
+
+use crate::activation::scale;
+use crate::gemm::batched_matmul;
+use crate::softmax::softmax;
+
+unsafe extern "C" {
+    fn launch_causal_mask_f32(scores: *mut c_void, batch: i32, rows: i32, cols: i32,
+                              offset: i32, stream: *mut c_void);
+    fn launch_causal_mask_bf16(scores: *mut c_void, batch: i32, rows: i32, cols: i32,
+                               offset: i32, stream: *mut c_void);
+}
+
+fn apply_causal_mask(scores: &Tensor, offset: usize) {
+    let ndim = scores.ndim();
+    let rows = scores.shape()[ndim - 2];
+    let cols = scores.shape()[ndim - 1];
+    let batch: usize = scores.shape()[..ndim - 2].iter().product();
+
+    unsafe {
+        match scores.dtype() {
+            DType::F32 => launch_causal_mask_f32(
+                scores.data_ptr() as *mut c_void,
+                batch as i32, rows as i32, cols as i32, offset as i32,
+                std::ptr::null_mut(),
+            ),
+            DType::BF16 => launch_causal_mask_bf16(
+                scores.data_ptr() as *mut c_void,
+                batch as i32, rows as i32, cols as i32, offset as i32,
+                std::ptr::null_mut(),
+            ),
+            _ => panic!("unsupported dtype for causal mask"),
+        }
+    }
+    xserv_cuda::device::synchronize().unwrap();
+}
+
+/// Multi-head attention (naive, materializes S×S score matrix).
+///
+/// q, k, v: [batch, num_heads, seq_len, head_dim] — contiguous, on GPU
+/// Returns: [batch, num_heads, seq_len, head_dim]
+pub fn attention(q: &Tensor, k: &Tensor, v: &Tensor, causal: bool) -> Tensor {
+    assert_eq!(q.ndim(), 4);
+    assert_eq!(k.ndim(), 4);
+    assert_eq!(v.ndim(), 4);
+    assert!(q.is_contiguous() && k.is_contiguous() && v.is_contiguous());
+
+    let batch = q.shape()[0];
+    let num_heads = q.shape()[1];
+    let q_len = q.shape()[2];
+    let head_dim = q.shape()[3];
+    let kv_len = k.shape()[2];
+
+    assert_eq!(k.shape(), &[batch, num_heads, kv_len, head_dim]);
+    assert_eq!(v.shape(), &[batch, num_heads, kv_len, head_dim]);
+
+    // scores = Q @ K^T → [B, H, q_len, kv_len]
+    let k_t = k.transpose(2, 3).contiguous();
+    let scores = batched_matmul(q, &k_t);
+
+    // Scale by 1/sqrt(head_dim)
+    let scale_factor = 1.0 / (head_dim as f32).sqrt();
+    let scaled_scores = scale(&scores, scale_factor);
+
+    // Causal mask
+    if causal {
+        let offset = kv_len - q_len;
+        apply_causal_mask(&scaled_scores, offset);
+    }
+
+    // Softmax
+    let weights = softmax(&scaled_scores);
+
+    // output = weights @ V → [B, H, q_len, head_dim]
+    batched_matmul(&weights, v)
+}