-
Notifications
You must be signed in to change notification settings - Fork 39
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
[Docs] Enable Llama3 inference bkc. (#2725)
- Loading branch information
Showing
4 changed files
with
222 additions
and
0 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,100 @@ | ||
# Llama3 Inference Best Known Method for Intel-Extension-For-Tensorflow on Intel GPU | ||
|
||
## Introduction | ||
Llama 3 is a collection of pretrained and fine-tuned generative text models ranging in scale from 8 billion to 70 billion parameters. For more detail information, please refer to [llama-3/keras](https://www.kaggle.com/models/metaresearch/llama-3/keras). | ||
|
||
This example shows how to run Llama3 8b inference with Intel® Extension for TensorFlow* on Intel GPU. | ||
|
||
## Hardware Requirements | ||
|
||
Verified Hardware Platforms: | ||
- Intel® Data Center GPU Max Series | ||
- Intel® Data Center GPU Flex Series 170 | ||
|
||
## Prerequisites | ||
### Dataset | ||
Follow [llama-3/keras/llama3_8b_en](https://www.kaggle.com/models/metaresearch/llama-3/keras/llama3_8b_en) to apply access permission and then download datasets. | ||
``` | ||
mkdir -p llama3_8b_en | ||
tar -xzvf llama3-keras-llama3_8b_en-v3.tar.gz -C ./llama3_8b_en | ||
``` | ||
|
||
### Prepare for GPU | ||
|
||
Refer to [Prepare](../common_guide_running.md#prepare) | ||
|
||
### Setup Running Environment | ||
* Setup for GPU | ||
```bash | ||
./pip_set_env.sh | ||
``` | ||
Note: This Llama3 keras3 implementation requires TensorFlow >= 2.16.1 and Intel® Extension for TensorFlow* >= 2.16.0.0. | ||
|
||
### Enable Running Environment | ||
|
||
Enable oneAPI running environment (only for GPU) and virtual running environment. | ||
|
||
* For GPU, refer to [Running](../common_guide_running.md#running) | ||
|
||
|
||
### Executes the Example with Python API | ||
#### Model Default Parameters | ||
| **Parameter** | **Default Value** | | ||
| :---: | :--- | | ||
| **model** | llama3_8b_en | | ||
| **dtype** | bfloat16 | | ||
| **data-dir** | ./ | | ||
| **input-tokens** | 32 | | ||
| **max-new-tokens** | 32 | | ||
| **num-beams** | 1 | | ||
| **num-iter** | 10 | | ||
| **num-warmup** | 3 | | ||
| **batch-size** | 1 | | ||
|
||
#### FP32 Inference | ||
``` | ||
python run_generate.py \ | ||
--model llama3_8b_en \ | ||
--dtype float32 \ | ||
--data-dir ./ \ | ||
--input-tokens 32 \ | ||
--max-new-tokens 32 | ||
``` | ||
|
||
#### BF16 Inference | ||
``` | ||
python run_generate.py \ | ||
--model llama3_8b_en \ | ||
--dtype bfloat16 \ | ||
--data-dir ./ \ | ||
--input-tokens 32 \ | ||
--max-new-tokens 32 | ||
``` | ||
|
||
## Example Output | ||
With successful execution, it will print out the following results: | ||
|
||
``` | ||
Prompt: Once upon a time, there existed a little girl, who liked to have adventures. She wanted to go to places and meet new people, and have fun. | ||
Iteration: 0, Time: xxx sec | ||
Iteration: 1, Time: xxx sec | ||
Iteration: 2, Time: xxx sec | ||
Iteration: 3, Time: xxx sec | ||
Iteration: 4, Time: xxx sec | ||
Iteration: 5, Time: xxx sec | ||
Iteration: 6, Time: xxx sec | ||
Iteration: 7, Time: xxx sec | ||
Iteration: 8, Time: xxx sec | ||
Iteration: 9, Time: xxx sec | ||
---------- Summary: ---------- | ||
Inference latency: xxx sec. | ||
Output: Once upon a time, there existed a little girl, who liked to have adventures. She wanted to go to places and meet new people, and have fun. She wanted to be a princess, and a pirate, and a fairy, and a mermaid, and a superhero, and a witch, and a queen. | ||
``` | ||
|
||
## FAQ | ||
|
||
1. If you get the following error log, refer to [Enable Running Environment](#Enable-Running-Environment) to Enable oneAPI running environment. | ||
``` | ||
tensorflow.python.framework.errors_impl.NotFoundError: libmkl_sycl.so.2: cannot open shared object file: No such file or directory | ||
``` |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,28 @@ | ||
#!/bin/bash | ||
|
||
# | ||
# Copyright (c) 2024 Intel Corporation | ||
# | ||
# 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. | ||
# | ||
|
||
ENV_NAME=env_itex | ||
deactivate | ||
rm -rf $ENV_NAME | ||
python -m venv $ENV_NAME | ||
source $ENV_NAME/bin/activate | ||
pip install --upgrade pip | ||
pip install tensorflow==2.16.1 | ||
pip install keras_nlp | ||
pip install numpy | ||
pip install --upgrade intel-extension-for-tensorflow-weekly[xpu] -f https://developer.intel.com/itex-whl-weekly |
Large diffs are not rendered by default.
Oops, something went wrong.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,93 @@ | ||
import os | ||
import tensorflow as tf | ||
import argparse | ||
import json | ||
import time | ||
import keras | ||
import keras_nlp | ||
import numpy as np | ||
import kagglehub | ||
|
||
# Download latest version | ||
#path = kagglehub.model_download("keras/llama3/keras/llama3_8b_en") | ||
#print("Path to model files:", path) | ||
|
||
parser = argparse.ArgumentParser() | ||
parser.add_argument( | ||
"--model", | ||
type=str, | ||
choices=["llama3_8b_en", "llama3_instruct_8b_en"], | ||
default="llama3_8b_en", | ||
help="the mdoel name", | ||
) | ||
parser.add_argument( | ||
"--data-dir", | ||
type=str, | ||
default="./", | ||
help="the dataset path", | ||
) | ||
parser.add_argument( | ||
"--dtype", | ||
type=str, | ||
choices=["float32", "bfloat16"], | ||
default="float32", | ||
help="float32, bfloat16" | ||
) | ||
parser.add_argument( | ||
"--prompt", default=None, type=str, help="input prompt for self-defined if needed" | ||
) | ||
parser.add_argument( | ||
"--input-tokens", | ||
default=None, | ||
choices=["32", "64", "128", "256", "512", "1024", "2016", "2017", "2048", "4096", "8192"], | ||
type=str, | ||
help="input tokens length if needed from prompt.json", | ||
) | ||
parser.add_argument( | ||
"--max-new-tokens", default=32, type=int, help="output max new tokens" | ||
) | ||
parser.add_argument("--num-beams", default=1, type=int, help="beam width") | ||
parser.add_argument("--num-iter", default=10, type=int, help="num iter") | ||
parser.add_argument("--num-warmup", default=3, type=int, help="num warmup") | ||
parser.add_argument("--batch-size", default=1, type=int, help="batch size") | ||
|
||
args = parser.parse_args() | ||
path = args.data_dir + args.model | ||
print("Dataset dir is: %s" % path) | ||
|
||
if args.dtype == "bfloat16": | ||
keras.config.set_floatx("bfloat16") | ||
model = keras_nlp.models.Llama3CausalLM.from_preset(path, dtype=args.dtype) | ||
if args.num_beams > 1: | ||
from keras_nlp.samplers import BeamSampler | ||
model.compile(sampler=BeamSampler(num_beams=args.num_beams)) | ||
else: | ||
model.compile(sampler="greedy") | ||
|
||
if args.prompt is not None: | ||
prompt = args.prompt | ||
elif args.input_tokens is not None: | ||
current_path = os.path.dirname(__file__) | ||
with open(str(current_path) + "/prompt.json") as f: | ||
prompt_pool = json.load(f) | ||
prompt = prompt_pool[args.input_tokens] | ||
print("Prompt: %s!" % prompt) | ||
|
||
total_time = 0.0 | ||
num_iter = args.num_iter | ||
num_warmup = args.num_warmup | ||
prompt = [prompt] * args.batch_size | ||
for i in range(num_iter): | ||
tic = time.time() | ||
output = model.generate( | ||
prompt, max_length=int(args.max_new_tokens)+int(args.input_tokens) | ||
) | ||
toc = time.time() | ||
print("Iteration: %d, Time: %.6f sec" % (i, toc - tic), flush=True) | ||
if i >= num_warmup: | ||
total_time += toc - tic | ||
|
||
print("\n", "-" * 10, "Summary:", "-" * 10) | ||
latency = total_time / (num_iter - num_warmup) | ||
print("Inference latency: %.3f sec." % latency) | ||
print("Output: %s." % output) |