while (off < tot) {
            int fragSize = tot - off;
            if (24 + fragSize > this.max_xmit) {
                // need fragementation
                msg.flags &= ~DCERPC_LAST_FRAG;
                fragSize = this.max_xmit - 24;
            } else {
                msg.flags |= DCERPC_LAST_FRAG;
                msg.alloc_hint = fragSize;
            }

            msg.length = 24 + fragSize;

# Note that linux cpu and cuda builds share the same toolchain now.
build:release_gpu_linux --config=cuda_clang_official
# Local test jobs has to be 4 because parallel_gpu_execute is fragile, I think
test:release_gpu_linux --test_timeout=300,450,1200,3600 --local_test_jobs=4 --run_under=//tensorflow/tools/ci_build/gpu_build:parallel_gpu_execute

# Deprecated release CPU config with non-hermetic toolchains.