如何用ppl 跑internlm2-20b模型的性能，也是使用pplnn-build/tools/benchmark_llama么？

There's no python binding related code in repository. Since ppl.nn.llm is ppl.nn, how can I integrate ppl.nn.llm with ppl.nn python binding ? Is it possible to use python to build a LLM chat interaction (or text completion) ?

使用命令： OMP_NUM_THREADS=1 torchrun --nproc_per_node
Export.py --ckpt_dir <llama_dir>
--tokenizer_path <llama_tokenizer_dir>/tokenizer.model
--export_path <path_to_store_exported_llama7b>
--fused_qkv 1 --fused_kvcache 1 --auto_causal 1
--quantized_cache 1 --dynamic_batching 1
export 模型的时候，报一些warning信息，这是什么原因？会有什么影响？需要解决么？谢谢
[W shape_type_inference.cpp:1920] Warning: The shape inference of pmx::SwiGLU type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding it in symbolic function. (function UpdateReliable)
[W shape_type_inference.cpp:1920] Warning: The shape inference of pmx::RowParallelLinear type is missing, so it may result in wrong shape inference for the exported graph. Please consider adding it in symbolic function. (function UpdateReliable)

How to set the output length for the llama7B benchmark script? It seems that the script is only profiing for single token generation.

看到代码里边有一些moe相关的算子。但是没看到moe模型的示例