Support internlm2

[RunningLeon]

This PR supports the conversion of internlm2 from hf to trt-llm checkpoints with :

• fp16/bf16
• tp

[PaulX1029]

做了相应修改后engine build不成功，请问可以怎样解决呢？
我的命令是：
python3 convert_checkpoint.py --model_dir /mnt/checkpoint/models/internlm2-chat-20b/ --dtype float16 --output_dir /mnt/tllm_checkpoint/internlm2-chat-20b/tllm_checkpoint_1gpu_tp1

trtllm-build --checkpoint_dir /mnt/tllm_checkpoint/internlm2-chat-20b/tllm_checkpoint_1gpu_tp1/ --output_dir /mnt/trt_engines/internlm2-chat-20b/fp16/1-gpu/ --gemm_plugin float16 --max_batch_size=32 --max_input_len=2048 --max_output_len=2048

[PaulX1029]

做了相应修改后engine build不成功，请问可以怎样解决呢？ 我的命令是： python3 convert_checkpoint.py --model_dir /mnt/checkpoint/models/internlm2-chat-20b/ --dtype float16 --output_dir /mnt/tllm_checkpoint/internlm2-chat-20b/tllm_checkpoint_1gpu_tp1

trtllm-build --checkpoint_dir /mnt/tllm_checkpoint/internlm2-chat-20b/tllm_checkpoint_1gpu_tp1/ --output_dir /mnt/trt_engines/internlm2-chat-20b/fp16/1-gpu/ --gemm_plugin float16 --max_batch_size=32 --max_input_len=2048 --max_output_len=2048
版本信息：

[RunningLeon]

@PaulX1029 Hi, have you rebuilt and reinstalled tensorrt-llm? You can find the installation location by pip3 show tensorrt_llmand check tensorrt_llm/models/__init__.py to see if MODEL_MAP is as expected:

[PaulX1029]

@RunningLeon 请问您是采用什么build的方式，我是从pip安装的trtllm，我想要跟您对齐build方式重新进行

[RunningLeon]

@RunningLeon 请问您是采用什么build的方式，我是从pip安装的trtllm，我想要跟您对齐build方式重新进行

#266 (comment)

[cqy930325]

@RunningLeon 很感谢你的工作！想问下internvl-1.5的internlm2-20b 网络跟普通internlm2-20b有什么区别吗，我用了PR里的转换脚本转出来之后都是乱码。

[ChengYouFancy]

@RunningLeon hi，we use lora finetuned the internlm2 model. Now we can convert the base model to llama, but not lora part, we tried to change the code of InternLM/tools/convert2llame.py to transfer lora to llama style, but did not work. Is there any other tools could work for lora?
tips: we can't export base model and lora to one single model, because we want to use https://github.com/NVIDIA/TensorRT-LLM/tree/main/examples/llama#run-llama-with-several-lora-checkpoints this feature. Thus, if we can transfer internlm2 lora to llama style, then we can use examples/hf_lora_convert.py to build trt-llm

[RunningLeon]

@nv-guomingz Hi, sorry to bother you, but when will this PR be merged? Do I need to fix the conflicts?
THX

[RunningLeon]

@nv-guomingz hi, the conflicts with main branch are resolved. Looking forward to your review comments. THX.

[DefTruth]

@RunningLeon 请问下为什么internlm2需要单独一个convert_checkpoint.py呢？而不是复用llama的convert_checkpoint.py，internlm是直接使用llama的convert_checkpoint.py

[RunningLeon]

@RunningLeon 请问下为什么internlm2需要单独一个convert_checkpoint.py呢？而不是复用llama的convert_checkpoint.py，internlm是直接使用llama的convert_checkpoint.py

hi, internlm2 W_qkv是在一起的，其次一些参数命名是和llama没有对齐的。因而没法直接使用llama的convert_checkpoint.py

[DefTruth]

@RunningLeon 请问下为什么internlm2需要单独一个convert_checkpoint.py呢？而不是复用llama的convert_checkpoint.py，internlm是直接使用llama的convert_checkpoint.py

hi, internlm2 W_qkv是在一起的，其次一些参数命名是和llama没有对齐的。因而没法直接使用llama的convert_checkpoint.py

Thank you for this explanation!

[nv-guomingz]

Hi @RunningLeon sorry for late response due to internal task priority.
Would u please rebase the code firstly and I'll try to merge your MR into main branch this week.

[RunningLeon]

Hi @RunningLeon sorry for late response due to internal task priority. Would u please rebase the code firstly and I'll try to merge your MR into main branch this week.

@nv-guomingz Done. Hope merging with main is OK.

[nv-guomingz]

Hi @RunningLeon sorry for late response due to internal task priority. Would u please rebase the code firstly and I'll try to merge your MR into main branch this week.

@nv-guomingz Done. Hope merging with main is OK.

Thanks @RunningLeon. Could u please rebase your commits into one single commit? That would be more easy for further integration.