TensorRT-LLMs/examples/models/contrib/chatglm2-6b/README.md

ChatGLM

This document explains how to build the ChatGLM2-6B, ChatGLM2-6B-32k models using TensorRT-LLM and run on a single GPU, a single node with multiple GPUs or multiple nodes with multiple GPUs.

• ChatGLM
  • Overview
  • Support Matrix
  • Model comparison
  • Tokenizer and special tokens comparison
  • Usage
    • 1. Download repo and weights from HuggingFace Transformers
    • 2. Convert weights from HF Transformers to TensorRT-LLM format
    • 3. Build TensorRT engine(s)
      • Enable plugins
      • In-flight batching
    • 4. Run inference
      • Single node, single GPU
      • Single node, multi GPU
    • 5. Run summarization task
    • Weight Only quantization
    • Smooth Quantization (SQ)
    • Activation-aware Weight Quantization (AWQ)
    • FP8 Quantization
  • Benchmark

Overview

The TensorRT-LLM ChatGLM implementation can be found in tensorrt_llm/models/chatglm/model.py. The TensorRT-LLM ChatGLM example code is located in examples/models/contrib/chatglm2-6b. There is one main file:

• examples/glm-4-9b/convert_checkpoint.py to convert a checkpoint from the HuggingFace (HF) Transformers format to the TensorRT-LLM format.

In addition, there are two shared files in the parent folder examples for inference and evaluation:

• ../../../run.py to run the inference on an input text;
• ../../../summarize.py to summarize the articles in the cnn_dailymail dataset.

Support Matrix

Model Name	FP16	FMHA	WO	SQ	AWQ	FP8	TP	PP	ST	C++	benchmark	IFB
chatglm2_6b	Y	Y	Y	Y	Y	Y	Y		Y	Y	Y	Y
chatglm2_6b_32k	Y	Y	Y		Y	Y	Y		Y	Y	Y	Y

• Model Name: the name of the model, the same as the name on HuggingFace
• FMHA: Fused MultiHead Attention (see introduction below)
• WO: Weight Only Quantization (int8 / int4)
• SQ: Smooth Quantization (int8)
• AWQ: Activation Aware Weight Quantization (int4)
• FP8: FP8 Quantization
• TP: Tensor Parallel
• PP: Pipeline Parallel
• ST: Strongly Typed
• C++: C++ Runtime
• benchmark: benchmark by python / C++ Runtime
• IFB: In-flight Batching (see introduction below)

Model comparison

Name	nL	nAH	nKH	nHW	nH	nF	nMSL	nV	bP2D	bBQKV	bBDense	Comments
chatglm2_6b	28	32	2	128	4096	13696	32768	65024	N	Y	N	Multi_query_attention, RMSNorm rather than LayerNorm in chatglm_6b
chatglm2_6b_32k	28	32	2	128	4096	13696	32768	65024	N	Y	N	RoPE base=160000 rather than 10000 in chatglm2_6b

• nL: number of layers
• nAH: number of attention heads
• nKH: number of kv heads (less than nAH if multi_query_attention is used)
• nHW: head width
• nH: hidden size
• nF: FFN hidden size
• nMSL: max sequence length (input + output)
• nV: vocabulary size
• bP2D: use position_encoding_2d (Y: Yes, N: No)
• bBQKV: use bias for QKV multiplication in self-attention
• bBDense: use bias for Dense multiplication in self-attention

Tokenizer and special tokens comparison

Name	Tokenizer	bos	eos	pad	cls	startofpiece	endofpiece	mask	smask	gmask
chatglm2_6b	ChatGLMTokenizer	1	2	0
chatglm2_6b_32k	ChatGLMTokenizer	1	2	0

Usage

The next section describe how to build the engine and run the inference demo.

1. Download repo and weights from HuggingFace Transformers

pip install -r requirements.txt
apt-get update
apt-get install git-lfs
rm -rf chatglm*

# clone one or more models we want to build
git clone https://huggingface.co/THUDM/chatglm2-6b      chatglm2_6b
git clone https://huggingface.co/THUDM/chatglm2-6b-32k  chatglm2_6b_32k

For more example codes, please refer to the examples/glm-4-9b/README.md.