f16 FA mask: allocate as F16 when flash_attn enabled

From upstream PR #23764. Avoids ggml_cast in graph builder.
Our kernel already reads masks as half — no changes needed.

Author: MrTrent

src/llama-graph.cpp

@@ -30,7 +30,10 @@ static ggml_tensor * build_attn_inp_kq_mask(
     const auto n_tokens = ubatch.n_tokens;
     const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
 
-    ggml_tensor * res = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
+    // flash attention requires an f16 mask
+    const auto type = cparams.flash_attn ? GGML_TYPE_F16 : GGML_TYPE_F32;
+
+    ggml_tensor * res = ggml_new_tensor_4d(ctx, type, n_kv, n_tokens/n_stream, 1, n_stream);
     ggml_set_input(res);
     ggml_set_name(res, "attn_inp_kq_mask");