ipex-llm/python/llm/example/transformers/native_int4

BigDL-LLM Native INT4 Inference Pipeline for Large Language Model

In this example, we show a pipeline to convert a large language model to BigDL-LLM native INT4 format, and then run inference on the converted INT4 model.

Note

: BigDL-LLM native INT4 format currently supports model family LLaMA(such as Vicuna, Guanaco, Koala, Baize, WizardLM, etc.), GPT-NeoX(such as RedPajama), BLOOM (such as Phoenix) and StarCoder.

Prepare Environment

We suggest using conda to manage environment:

conda create -n llm python=3.9
conda activate llm

pip install --pre --upgrade bigdl-llm[all]

Run Example

python ./native_int4_pipeline.py --thread-num THREAD_NUM --model-family MODEL_FAMILY --repo-id-or-model-path MODEL_PATH

arguments info:

• --thread-num THREAD_NUM: required argument defining the number of threads to use for inference. It is default to be 2.

• --model-family MODEL_FAMILY: required argument defining the model family of the large language model (supported option: 'llama', 'gptneox', 'bloom', 'starcoder'). It is default to be 'llama'.

• --repo-id-or-model-path MODEL_PATH: required argument defining the path to the huggingface checkpoint folder for the model.

  Note

  MODEL_PATH should fits your inputed MODEL_FAMILY.

• --promp PROMPT: optional argument defining the prompt to be infered. It is default to be 'Q: What is CPU? A:'.

• --tmp-path TMP_PATH: optional argument defining the path to store intermediate model during the conversion process. It is default to be '/tmp'.

Sample Output for Inference

Model family LLaMA

--------------------  bigdl-llm based tokenizer  --------------------
Inference time: xxxx s
Output:
[' It stands for Central Processing Unit. It’s the part of your computer that does the actual computing, or calculating. The first computers were all about adding machines']
--------------------  HuggingFace transformers tokenizer  --------------------
Please note that the loading of HuggingFace transformers tokenizer may take some time.

Inference time: xxxx s
Output:
['Central Processing Unit (CPU) is the main component of a computer system, also known as microprocessor. It executes the instructions of software programmes (also']
--------------------  fast forward  --------------------

bigdl-llm timings:        load time =    xxxx ms
bigdl-llm timings:      sample time =    xxxx ms /    32 runs   (    xxxx ms per token)
bigdl-llm timings: prompt eval time =    xxxx ms /     9 tokens (    xxxx ms per token)
bigdl-llm timings:        eval time =    xxxx ms /    31 runs   (    xxxx ms per token)
bigdl-llm timings:       total time =    xxxx ms
Inference time (fast forward): xxxx s
Output:
{'id': 'cmpl-c87e5562-281a-4837-8665-7b122948e0e8', 'object': 'text_completion', 'created': 1688368515, 'model': './bigdl_llm_llama_q4_0.bin', 'choices': [{'text': ' CPU stands for Central Processing Unit. This means that the processors in your computer are what make it run, so if you have a Pentium 4', 'index': 0, 'logprobs': None, 'finish_reason': 'length'}], 'usage': {'prompt_tokens': 9, 'completion_tokens': 32, 'total_tokens': 41}}

Model family GPT-NeoX

--------------------  bigdl-llm based tokenizer  --------------------
Inference time: xxxx s
Output:
[' Central processing unit, also known as processor, is a specialized microchip designed to execute all the instructions of computer programs rapidly and efficiently. Most personal computers have one or']
--------------------  HuggingFace transformers tokenizer  --------------------
Please note that the loading of HuggingFace transformers tokenizer may take some time.

Inference time: xxxx s
Output:
[' The Central Processing Unit, or CPU, is the component of a computer that executes all instructions for carrying out different functions. It is the brains of the operation, and']
--------------------  fast forward  --------------------
Gptneox.generate: prefix-match hit

gptneox_print_timings:        load time =     xxxx ms
gptneox_print_timings:      sample time =     xxxx ms /    32 runs   (    xxxx ms per run)
gptneox_print_timings: prompt eval time =     xxxx ms /     8 tokens (    xxxx ms per token)
gptneox_print_timings:        eval time =     xxxx ms /    31 runs   (    xxxx ms per run)
gptneox_print_timings:       total time =     xxxx ms
Inference time (fast forward): xxxx s
Output:
{'id': 'cmpl-a20fc4a1-3a00-4e77-a6cf-0dd0da6b9a59', 'object': 'text_completion', 'created': 1686557799, 'model': './bigdl_llm_gptneox_q4_0.bin', 'choices': [{'text': ' Core Processing Unit  or Central Processing Unit  is the brain of your computer, system software runs on it and handles all important tasks in your computer. i', 'index': 0, 'logprobs': None, 'finish_reason': 'length'}], 'usage': {'prompt_tokens': 9, 'completion_tokens': 32, 'total_tokens': 41}}

Model family BLOOM

--------------------  bigdl-llm based tokenizer  --------------------
Inference time: xxxx s
Output:
[' Central Processing Unit</s>The present invention relates to a method of manufacturing an LED device, and more particularly to the manufacture of high-powered LED devices. The inventive']
--------------------  HuggingFace transformers tokenizer  --------------------
Please note that the loading of HuggingFace transformers tokenizer may take some time.

Inference time: xxxx s
Output:
[' Central Processing Unit</s>The present invention relates to a method of manufacturing an LED device, and more particularly to the manufacture of high-powered LED devices. The inventive']
--------------------  fast forward  --------------------


inference:    mem per token = 24471324 bytes
inference:      sample time =     xxxx ms
inference: evel prompt time =     xxxx ms / 1 tokens / xxxx ms per token
inference:     predict time =     xxxx ms / 4 tokens / xxxx ms per token
inference:       total time =     xxxx ms
Inference time (fast forward): xxxx s
Output:
{'id': 'cmpl-4ec29030-f0c4-43d6-80b0-5f5fb76c169d', 'object': 'text_completion', 'created': 1687852341, 'model': './bigdl_llm_bloom_q4_0.bin', 'choices': [{'text': ' the Central Processing Unit</s>', 'index': 0, 'logprobs': None, 'finish_reason': None}], 'usage': {'prompt_tokens': 6, 'completion_tokens': 5, 'total_tokens': 11}}

Model family StarCoder

--------------------  bigdl-llm based tokenizer  --------------------
Inference time: xxxx s
Output:
[' 2.56 GHz, 2.56 GHz, 2.56 GHz, 2.56 GHz, ']
--------------------  HuggingFace transformers tokenizer  --------------------
Please note that the loading of HuggingFace transformers tokenizer may take some time.

Inference time: xxxx s
Output:
[' 2.56 GHz, 2.56 GHz, 2.56 GHz, 2.56 GHz, ']
--------------------  fast forward  --------------------


bigdl-llm:    mem per token =   313720 bytes
bigdl-llm:      sample time =     xxxx ms
bigdl-llm: evel prompt time =     xxxx ms
bigdl-llm:     predict time =     xxxx ms / 31 tokens / xxxx ms per token
bigdl-llm:       total time =     xxxx ms
Inference time (fast forward): xxxx s
Output:
{'id': 'cmpl-72bc4d13-d8c9-4bcb-b3f4-50a69863d534', 'object': 'text_completion', 'created': 1687852580, 'model': './bigdl_llm_starcoder_q4_0.bin', 'choices': [{'text': ' 0.50, B: 0.25, C: 0.125, D: 0.0625', 'index': 0, 'logprobs': None, 'finish_reason': None}], 'usage': {'prompt_tokens': 8, 'completion_tokens': 32, 'total_tokens': 40}}