ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

torch 2.3 inference docker (#12517)

Browse source 
* torch 2.3 inference docker

* Update README.md

* add convert code

* rename image

* remove 2.1 and add graph example

* Update README.md
This commit is contained in:
Heyang Sun 2024-12-13 10:47:04 +08:00 committed by GitHub
parent b747f3f6b8
commit fa261b8af1
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 352 additions and 6 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

25
docker/llm/inference/xpu/docker/Dockerfile
View file

@ -1,4 +1,4 @@
FROM intel/oneapi-basekit:2024.0.1-devel-ubuntu22.04
FROM intel/oneapi:2024.2.1-0-devel-ubuntu22.04
ARG http_proxy
ARG https_proxy
@ -6,7 +6,7 @@ ARG https_proxy
ENV TZ=Asia/Shanghai
ENV PYTHONUNBUFFERED=1
# When cache is enabled SYCL runtime will try to cache and reuse JIT-compiled binaries.
# When cache is enabled SYCL runtime will try to cache and reuse JIT-compiled binaries.
ENV SYCL_CACHE_PERSISTENT=1
COPY chat.py /llm/chat.py
@ -29,6 +29,19 @@ RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRO
env DEBIAN_FRONTEND=noninteractive apt-get update && \
# add-apt-repository requires gnupg, gpg-agent, software-properties-common
apt-get install -y --no-install-recommends gnupg gpg-agent software-properties-common && \
export PRE_DIR=$(pwd) && \
# Install Compute Runtime
mkdir -p /tmp/neo && \
cd /tmp/neo && \
wget https://github.com/oneapi-src/level-zero/releases/download/v1.18.5/level-zero_1.18.5+u22.04_amd64.deb && \
wget https://github.com/intel/intel-graphics-compiler/releases/download/igc-1.0.17791.9/intel-igc-core_1.0.17791.9_amd64.deb && \
wget https://github.com/intel/intel-graphics-compiler/releases/download/igc-1.0.17791.9/intel-igc-opencl_1.0.17791.9_amd64.deb && \
wget https://github.com/intel/compute-runtime/releases/download/24.39.31294.12/intel-level-zero-gpu_1.6.31294.12_amd64.deb && \
wget https://github.com/intel/compute-runtime/releases/download/24.39.31294.12/intel-opencl-icd_24.39.31294.12_amd64.deb && \
wget https://github.com/intel/compute-runtime/releases/download/24.39.31294.12/libigdgmm12_22.5.2_amd64.deb && \
dpkg -i *.deb && \
rm -rf /tmp/neo && \
cd $PRE_DIR && \
# Add Python 3.11 PPA repository
add-apt-repository ppa:deadsnakes/ppa -y && \
apt-get install -y --no-install-recommends python3.11 git curl wget && \
@ -41,13 +54,12 @@ RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRO
python3 get-pip.py && \
rm get-pip.py && \
pip install --upgrade requests argparse urllib3 && \
pip install --pre --upgrade ipex-llm[xpu] --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/ && \
pip install --pre --upgrade ipex-llm[xpu_arc] --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/ && \
pip install --pre pytorch-triton-xpu==3.0.0+1b2f15840e --index-url https://download.pytorch.org/whl/nightly/xpu && \
# Fix Trivy CVE Issues
pip install transformers==4.36.2 && \
pip install transformers_stream_generator einops tiktoken && \
# Install opencl-related repos
apt-get update && \
apt-get install -y --no-install-recommends intel-opencl-icd=23.35.27191.42-775~22.04 intel-level-zero-gpu=1.3.27191.42-775~22.04 level-zero=1.14.0-744~22.04 && \
# Install related libary of chat.py
pip install --upgrade colorama && \
# Download all-in-one benchmark and examples
@ -71,7 +83,7 @@ RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRO
# Download Deepspeed-AutoTP
cp -r ./ipex-llm/python/llm/example/GPU/Deepspeed-AutoTP/ ./Deepspeed-AutoTP && \
# Install related library of Deepspeed-AutoTP
pip install oneccl_bind_pt==2.1.100 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/ && \
pip install oneccl_bind_pt==2.3.100 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/ && \
pip install git+https://github.com/microsoft/DeepSpeed.git@ed8aed5 && \
pip install git+https://github.com/intel/intel-extension-for-deepspeed.git@0eb734b && \
pip install mpi4py && \
@ -82,3 +94,4 @@ RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRO
WORKDIR /llm/
ENV BIGDL_CHECK_DUPLICATE_IMPORT=0

48
python/llm/example/GPU/GraphMode/README.md Normal file
View file

@ -0,0 +1,48 @@
# Torch Graph Mode
Here, we provide how to run [torch graph mode](https://pytorch.org/blog/optimizing-production-pytorch-performance-with-graph-transformations/) on Intel Arc™ A-Series Graphics with ipex-llm, and [gpt2-medium](https://huggingface.co/openai-community/gpt2-medium) for classification task is used as illustration:
### 1. Install
```bash
conda create -n ipex-llm python=3.11
conda activate ipex-llm
pip install --pre --upgrade ipex-llm[xpu_arc] --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/cn/
pip install --pre pytorch-triton-xpu==3.0.0+1b2f15840e --index-url https://download.pytorch.org/whl/nightly/xpu
conda install -c conda-forge libstdcxx-ng
unset OCL_ICD_VENDORS
```
### 2. Configures OneAPI environment variables
> [!NOTE]
> Skip this step if you are running on Windows.
This is a required step on Linux for APT or offline installed oneAPI. Skip this step for PIP-installed oneAPI.
```bash
source /opt/intel/oneapi/setvars.sh
```
### 3. Run
Convert text-generating GPT2-Medium to the classification:
```bash
# The convert step needs to access the internet
export http_proxy=http://your_proxy_url
export https_proxy=http://your_proxy_url
# This will yield gpt2-medium-classification under /llm/models in the container
python convert-model-textgen-to-classfication.py --model-path MODEL_PATH
```
This will yield a mode directory ends with '-classification' neart your input model path.
Benchmark GPT2-Medium's performance with IPEX-LLM engine:
``` sbash
ipexrun xpu gpt2-graph-mode-benchmark.py --device xpu --engine ipex-llm --batch 16 --model-path MODEL_PATH
# You will see the key output like:
# Average time taken (excluding the first two loops): xxxx seconds, Classification per seconds is xxxx
```

57
python/llm/example/GPU/GraphMode/convert-model-textgen-to-classfication.py Normal file
View file

@ -0,0 +1,57 @@
#
# Copyright 2016 The BigDL Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# This is modified from https://github.com/intel-sandbox/customer-ai-test-code/blob/main/convert-model-textgen-to-classfication.py
#
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification, AutoConfig, AutoModelForCausalLM
import argparse
parser = argparse.ArgumentParser(description='Process some integers.')
parser.add_argument('--model_path', type=str, help='an string for the device')
args = parser.parse_args()
model_path = args.model_path
dtype=torch.bfloat16
num_labels = 5
model_name=model_path
save_directory = model_name + "-classification"
# Initialize the tokenizer
# Need padding from the left and padding to 1024
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
# tokenizer.padding_side = "left"
tokenizer.pad_token = tokenizer.eos_token
tokenizer.save_pretrained(save_directory)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=dtype, pad_token_id=tokenizer.eos_token_id,)
config = AutoConfig.from_pretrained(model_name)
print("text gen model")
print(model)
print(config)
model = AutoModelForSequenceClassification.from_pretrained(model_name, num_labels=num_labels, torch_dtype=dtype)
save_directory = model_name + "-classification"
model.save_pretrained(save_directory)
model = AutoModelForSequenceClassification.from_pretrained(save_directory, torch_dtype=dtype, pad_token_id=tokenizer.eos_token_id)
config = AutoConfig.from_pretrained(save_directory)
print("text classification model")
print(model)
print(config)

228
python/llm/example/GPU/GraphMode/gpt2-graph-mode-benchmark.py Normal file
View file

@ -0,0 +1,228 @@
#
# Copyright 2016 The BigDL Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# This is modified from https://github.com/intel-sandbox/customer-ai-test-code/blob/main/gpt2-benchmark-for-sangfor.py
#
import torch
import time
import argparse
from transformers import GPT2ForSequenceClassification, AutoTokenizer, AutoModelForSequenceClassification, AutoConfig, Qwen2ForSequenceClassification
from torch.profiler import profile, record_function, ProfilerActivity, schedule
# Get the batch size and device
parser = argparse.ArgumentParser(description='Process some integers.')
parser.add_argument('--batch_size', type=int, default=1, help='an integer for the batch size')
parser.add_argument('--device', type=str, default='cpu', help='an string for the device')
parser.add_argument('--profile', type=bool, default=False, help='enable protch profiler for CPU/XPU')
parser.add_argument('--engine', type=str, default='ipex-llm', help='an string for the device')
parser.add_argument('--model_path', type=str, help='an string for the device')
args = parser.parse_args()
enable_profile=args.profile
batch_size = args.batch_size
device = args.device
engine = args.engine
model_path = args.model_path
print(f"The batch size is: {batch_size}, device is {device}")
######################################################################################
# PyTorch Profiling with IPEX
# export IPEX_ZE_TRACING=1
# export ZE_ENABLE_TRACING_LAYER=1
import contextlib
def profiler_setup(profiling=False, *args, **kwargs):
if profiling:
return torch.profiler.profile(*args, **kwargs)
else:
return contextlib.nullcontext()
my_schedule = schedule(
skip_first=6,
wait=1,
warmup=1,
active=1
)
# also define a handler for outputing results
def trace_handler(p):
if(device == 'xpu'):
print(p.key_averages().table(sort_by="self_xpu_time_total", row_limit=20))
print(p.key_averages().table(sort_by="cpu_time_total", row_limit=20))
# p.export_chrome_trace("./trace_" + str(p.step_num) + ".json")
#######################################################################################
dtype = torch.bfloat16 if device == 'cpu' else torch.float16
num_labels = 5
model_name = model_path
model_name = model_name + "-classification"
model_name_ov = model_name + "-ov"
model_name_ov = model_name_ov + "-fp16"
if (engine == 'ipex') :
import torch
import intel_extension_for_pytorch as ipex
# Need padding from the left and padding to 1024
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
tokenizer.padding_side = "left"
tokenizer.pad_token = tokenizer.eos_token
model = AutoModelForSequenceClassification.from_pretrained(model_name, torch_dtype=dtype,
pad_token_id=tokenizer.eos_token_id,
low_cpu_mem_usage=True
).eval().to(device)
elif (engine == 'ipex-llm'):
from ipex_llm.transformers import AutoModelForSequenceClassification
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name,trust_remote_code=True)
tokenizer.padding_side = "left"
tokenizer.pad_token = tokenizer.eos_token
model = AutoModelForSequenceClassification.from_pretrained(model_name,
torch_dtype=dtype,
load_in_low_bit="fp16",
pad_token_id=tokenizer.eos_token_id,
low_cpu_mem_usage=True).to(device)
model = torch.compile(model, backend='inductor')
print(model)
else:
from optimum.intel import OVModelForSequenceClassification
tokenizer = AutoTokenizer.from_pretrained(model_name_ov, trust_remote_code=True)
tokenizer.padding_side = "left"
tokenizer.pad_token = tokenizer.eos_token
model = OVModelForSequenceClassification.from_pretrained(model_name_ov, torch_dtype=dtype).to(device)
# Intel(R) Extension for PyTorch*
if engine == 'ipex':
if device == 'cpu':
# model = ipex.llm.optimize(model, dtype=dtype, inplace=True, deployment_mode=True)
# ############## TorchDynamo ###############
model = ipex.optimize(model, dtype=torch.bfloat16, weights_prepack=False)
model = torch.compile(model, backend='ipex')
# ##########################################
else: # Intel XPU
#model = ipex.llm.optimize(model, dtype=dtype, device="xpu", inplace=True)
model = ipex.optimize(model, dtype=dtype, inplace=True)
model=torch.compile(model, backend="inductor")
print(model)
# # #######calulate the total num of parameters########
# def model_size(model):
# return sum(t.numel() for t in model.parameters())
# print(f"GPT2 size: {model_size(model)/1000**2:.1f}M parameters")
# # # #######print model information ###################
# print(model)
# ########Enable the BetterTransformer ###################
# only Better Transformer only support GPT2, not support Qwen2
# model = BetterTransformer.transform(model)
#elif engine == 'ipex-llm':
# model = ipex.optimize(model, dtype=dtype, inplace=True)
# model=torch.compile(model) #backend="inductor")
elif engine == 'ov':
print("OV inference")
prompt = ["this is the first prompt"]
prompts = prompt * batch_size
#print(prompts)
# Tokenize the batch of prompts
inputs = tokenizer(prompts, return_tensors="pt", padding="max_length", max_length=1024, truncation=True)
# print(inputs)
if engine == 'ipex' or engine == 'ipex-llm':
#ipex need move the inputs to device, but OV doesn't need
inputs.to(device)
# Initialize an empty list to store elapsed times
elapsed_times = []
# Loop for batch processing 10 times and calculate the time for every loop
with profiler_setup(profiling=enable_profile, activities=[ProfilerActivity.CPU, ProfilerActivity.XPU],
schedule=my_schedule,
on_trace_ready=trace_handler,
# on_trace_ready=torch.profiler.tensorboard_trace_handler('./log/gpt2'),
record_shapes=True,
with_stack=True
) as prof:
for i in range(10):
start_time = time.time()
# Perform inference
with torch.inference_mode():
# logits = model(**inputs).logits
outputs = model(**inputs)
logits = outputs.logits
# Get the predicted class for each input in the batch
predicted_class_ids = logits.argmax(dim=1).tolist()
end_time = time.time()
elapsed_time = end_time - start_time
# Save the elapsed time in the list
elapsed_times.append(elapsed_time)
if(enable_profile):
prof.step()
# print(outputs)
# print(type(outputs))
# print("logits.shape is " + str(logits.shape))
# print(logits)
# print(predicted_class_ids)
elif engine == 'ov':
print("OV inference")
# Initialize an empty list to store elapsed times
elapsed_times = []
# Loop for batch processing 10 times and calculate the time for every loop
for i in range(10):
start_time = time.time()
outputs = model(**inputs)
logits = outputs.logits
# Get the predicted class for each input in the batch
predicted_class_ids = logits.argmax(dim=1).tolist()
end_time = time.time()
elapsed_time = end_time - start_time
# Save the elapsed time in the list
elapsed_times.append(elapsed_time)
# print(outputs)
# print(type(outputs))
# print("logits.shape is " + str(logits.shape))
# print(logits)
# print(predictions)
#print(predicted_class_ids)
# Skip the first two values and calculate the average of the remaining elapsed times
average_elapsed_time = sum(elapsed_times[2:]) / len(elapsed_times[2:])
classfication_per_second = batch_size/average_elapsed_time
print(f"Average time taken (excluding the first two loops): {average_elapsed_time:.4f} seconds, Classification per seconds is {classfication_per_second:.4f}")