diff --git a/csrc/libtorch_stable/quantization/awq/gemm_kernels.cu b/csrc/libtorch_stable/quantization/awq/gemm_kernels.cu index c3702c52ef..43301f0fe2 100644 --- a/csrc/libtorch_stable/quantization/awq/gemm_kernels.cu +++ b/csrc/libtorch_stable/quantization/awq/gemm_kernels.cu @@ -528,5 +528,9 @@ torch::stable::Tensor awq_gemm(torch::stable::Tensor _in_feats, group_size, split_k_iters, in_feats, kernel, scaling_factors, zeros, num_in_feats, num_in_channels, num_out_channels, out_feats); } - return torch::stable::sum(_out_feats, 0); + // Pass the reduction dim via a named, lifetime-stable local rather than a + // bare `0`. + constexpr int64_t sum_dim = 0; + return torch::stable::sum( + _out_feats, torch::headeronly::IntHeaderOnlyArrayRef(&sum_dim, 1)); }