fix hang due to small rdma_chunk_size (#317)

Co-authored-by: zhiyi Hu <zhiyihu@U-NYQQMGK0-2250.local>

csrc/kernels/internode.cu

@@ -1851,6 +1851,7 @@ void combine(cudaDataType_t type,
     EP_HOST_ASSERT(num_forwarder_warps > NUM_MAX_NVL_PEERS and num_forwarder_warps % num_rdma_ranks == 0);
     EP_HOST_ASSERT(num_max_nvl_chunked_recv_tokens % num_rdma_ranks == 0);
     EP_HOST_ASSERT(num_max_nvl_chunked_recv_tokens / num_rdma_ranks > std::max(num_max_rdma_chunked_send_tokens, num_max_nvl_chunked_send_tokens));
+    EP_HOST_ASSERT(num_max_rdma_chunked_send_tokens >= num_warps_per_forwarder);
     EP_HOST_ASSERT(type == CUDA_R_16BF);
 
     SETUP_LAUNCH_CONFIG(num_channels * 2, (num_forwarder_warps + 1) * 32, stream);