CUDA: missing PDL sync for FWHT, better fallback (#23690)

2026-05-26 11:05:51 +08:00
parent 35c9b1f39e
commit 192d8ae8b8
3 changed files with 17 additions and 25 deletions
@@ -2596,9 +2596,7 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
    bool use_batched_cublas_f32  = src0->type == GGML_TYPE_F32;

    const int32_t hint = ggml_get_op_params_i32(dst, 1);
-    if (hint == GGML_HINT_SRC0_IS_HADAMARD) {
-        GGML_ASSERT(!split);
-        ggml_cuda_op_fwht(ctx, src1, dst);
+    if (hint == GGML_HINT_SRC0_IS_HADAMARD && !split && ggml_cuda_op_fwht(ctx, src1, dst)) {
        return;
    }