IPEX-LLM is a PyTorch library for running LLM on Intel CPU and GPU (e.g., local PC with iGPU, discrete GPU such as Arc, Flex and Max) with very low latency.
This example goes over how to use LlamaIndex to interact with ipex-llm for text generation and chat on Intel GPU.
Note
You could refer to here for full examples of IpexLLM. Please note that for running on Intel GPU, please specify -d 'xpu' or -d 'xpu:<device_id>' in command argument when running the examples.
After the prerequisites installation, you should have created a conda environment with all prerequisites installed, activate your conda environment and install llama-index-llms-ipex-llm as follows:
Setting device_map="xpu" when initializing IpexLLM will put the LLM model on Intel GPU and benefit from IPEX-LLM optimizations.
Note
If you have multiple Intel GPUs available, you could set device="xpu:<device_id>", in which device_id is counted from 0. device="xpu" is equal to device="xpu:0" by default.
Before loading the Zephyr model, you’ll need to define completion_to_prompt and messages_to_prompt for formatting prompts. Follow proper prompt format for zephyr-7b-alpha following the model card. This is essential for preparing inputs that the model can interpret accurately. Load the Zephyr model locally using IpexLLM using IpexLLM.from_model_id. It will load the model directly in its Huggingface format and convert it automatically to low-bit format for inference.
# Transform a string into input zephyr-specific input
# ensure we start with a system prompt, insert blank if needed
ifnot prompt.startswith("<|system|>\n"):
prompt ="<|system|>\n</s>\n"+ prompt
# add final assistant prompt
prompt = prompt +"<|assistant|>\n"
return prompt
from llama_index.llms.ipex_llm import IpexLLM
llm = IpexLLM.from_model_id(
model_name="HuggingFaceH4/zephyr-7b-alpha",
tokenizer_name="HuggingFaceH4/zephyr-7b-alpha",
context_window=512,
max_new_tokens=128,
generate_kwargs={"do_sample": False},
completion_to_prompt=completion_to_prompt,
messages_to_prompt=messages_to_prompt,
device_map="xpu",
)
Please note that in this example we’ll use HuggingFaceH4/zephyr-7b-alpha model for demostration. It requires updating transformers and tokenizers packages.
message =ChatMessage(role="user",content="What is AI?")
resp = llm.stream_chat([message],max_tokens=256)
for r in resp:
print(r.delta,end="")
Alternatively, you might save the low-bit model to disk once and use from_model_id_low_bit instead of from_model_id to reload it for later use - even across different machines. It is space-efficient, as the low-bit model demands significantly less disk space than the original model. And from_model_id_low_bit is also more efficient than from_model_id in terms of speed and memory usage, as it skips the model conversion step.
To save the low-bit model, use save_low_bit as follows. Then load the model from saved lowbit model path. Also use device_map to load the model to xpu.
Note that the saved path for the low-bit model only includes the model itself but not the tokenizers. If you wish to have everything in one place, you will need to manually download or copy the tokenizer files from the original model’s directory to the location where the low-bit model is saved.
Try stream completion using the loaded low-bit model.
saved_lowbit_model_path = (
"./zephyr-7b-alpha-low-bit"# path to save low-bit model
)
llm._model.save_low_bit(saved_lowbit_model_path)
del llm
llm_lowbit = IpexLLM.from_model_id_low_bit(
model_name=saved_lowbit_model_path,
tokenizer_name="HuggingFaceH4/zephyr-7b-alpha",
# tokenizer_name=saved_lowbit_model_path, # copy the tokenizers to saved path if you want to use it this way
context_window=512,
max_new_tokens=64,
completion_to_prompt=completion_to_prompt,
generate_kwargs={"do_sample": False},
device_map="xpu",
)
response_iter = llm_lowbit.stream_complete("What is Large Language Model?")
