Merge pull request #50 from purdue-aalp/huggingface_app

add huggingface sample app

Author: LAhmos

.gitignore

@@ -13,4 +13,6 @@ src/cuda/rodinia/3.1/cuda/nn/nn
 src/cuda/rodinia/3.1/cuda/particlefilter/particlefilter_float
 src/cuda/rodinia/3.1/cuda/particlefilter/particlefilter_naive
 src/cuda/rodinia/3.1/cuda/pathfinder/pathfinder
-4.2
+4.2
+.venv/
+__pycache__/

src/Makefile

@@ -4,7 +4,7 @@ $(error You must run "source setup_environment before calling make")
 endif
 
 ifeq ($(CUDA_GT_7), 1)
-all:  GPU_Microbenchmark microbench rodinia_2.0-ft cutlass rodinia-3.1 pannotia proxy-apps ispass-2009 lonestargpu-2.0 polybench custom_apps heterosync cuda_samples mlperf_inference vllm # mlperf_training
+all:  GPU_Microbenchmark microbench rodinia_2.0-ft cutlass rodinia-3.1 pannotia proxy-apps ispass-2009 lonestargpu-2.0 polybench custom_apps heterosync cuda_samples mlperf_inference vllm huggingface # mlperf_training
 else
 ifeq ($(CUDA_GT_4), 1)
 all:   pannotia rodinia_2.0-ft proxy-apps dragon-naive microbench rodinia-3.1 ispass-2009 dragon-cdp lonestargpu-2.0 polybench parboil shoc custom_apps
@@ -19,7 +19,7 @@ accelwattch_hw_power: rodinia-3.1_hw_power parboil_hw_power cuda_samples-11.0_hw
 #Disable clean for now, It has a bug!
 # clean_dragon-naive clean_pannotia clean_proxy-apps
 
-clean: clean_mlperf_inference clean_rodinia_2.0-ft clean_dragon-cdp  clean_ispass-2009 clean_lonestargpu-2.0 clean_custom_apps clean_parboil clean_cutlass clean_rodinia-3.1 clean_heterosync clean_UVMSmart_test clean_cuda_samples
+clean: clean_mlperf_inference clean_rodinia_2.0-ft clean_dragon-cdp  clean_ispass-2009 clean_lonestargpu-2.0 clean_custom_apps clean_parboil clean_cutlass clean_rodinia-3.1 clean_heterosync clean_UVMSmart_test clean_cuda_samples clean_huggingface
 clean_accelwattch: clean_rodinia-3.1 clean_parboil clean_cutlass clean_cuda_samples-11.0 clean_cuda_samples_hw_power clean_rodinia-3.1_hw_power clean_parboil_hw_power clean_accelwattch_ubench
 
 clean_data:
@@ -511,6 +511,14 @@ pytorch_examples:
 	echo "#!/bin/bash\npython $(BINDIR)/$(BINSUBDIR)/pytorch_examples/vae/main.py --epochs=1" > $(BINDIR)/$(BINSUBDIR)/inference_vae
 	chmod u+x $(BINDIR)/$(BINSUBDIR)/inference_vae
 
+huggingface:
+	mkdir -p $(BINDIR)/$(BINSUBDIR)/huggingface
+	cp -r cuda/huggingface $(BINDIR)/$(BINSUBDIR)
+	bash $(BINDIR)/$(BINSUBDIR)/huggingface/setup_environment.sh
+	echo "source $(BINDIR)/$(BINSUBDIR)/huggingface/setup_environment.sh && python3 $(BINDIR)/$(BINSUBDIR)/huggingface/helloworld.py" > $(BINDIR)/$(BINSUBDIR)/huggingface/helloworld
+	chmod u+x $(BINDIR)/$(BINSUBDIR)/huggingface/helloworld
+	chmod u+x $(BINDIR)/$(BINSUBDIR)/huggingface/*.py
+
 clean_heterosync:
 	rm -rf cuda/heterosync
 
@@ -685,3 +693,6 @@ clean_pytorch_examples:
 
 clean_cuda_samples:
 	make clean -C ./cuda/cuda-samples/build
+
+clean_huggingface:
+	rm -rf $(BINDIR)/$(BINSUBDIR)/huggingface

src/cuda/huggingface/README.md

@@ -0,0 +1,23 @@
+# Huggingface example
+
+Run huggingface models
+
+## Loading gated model
+
+Make sure to create a `.env` file in the folder with the following content:
+
+```bash
+HF_TOKEN=hf_...
+```
+
+## Setup environment
+
+```
+./setup_environment.sh
+```
+
+## Run example
+
+```bash
+./helloworld.py --model_name "openai-community/gpt2-large" --prompt "Hello World! In a galaxy far, far away..."
+```

src/cuda/huggingface/helloworld.py

@@ -0,0 +1,26 @@
+# Hello world for huggingface
+import argparse
+import transformers
+from dotenv import load_dotenv
+
+from util import getCommonParser
+
+def main():
+    load_dotenv()
+    
+    helloworldParser = argparse.ArgumentParser(parents=[getCommonParser()])
+    helloworldParser.add_argument("--prompt", type=str, default="Hello World! In a galaxy far, far away...")
+    helloworldParser.add_argument("--max_length", type=int, default=256)
+    args = helloworldParser.parse_args()
+    
+    helloworld_str = f"Running Hello world for \"{args.model_name}\" with prompt: \"{args.prompt}\""
+    print(helloworld_str)
+    print("=" * len(helloworld_str))
+    
+    pipe = transformers.pipeline(task="text-generation", model=args.model_name, device=args.device)
+    results = pipe(args.prompt, max_length=args.max_length, truncation=True)
+    for result in results:
+        print(result["generated_text"])
+
+if __name__ == "__main__":
+    main()

src/cuda/huggingface/requirements.txt

@@ -0,0 +1,7 @@
+# Huggingface
+transformers
+datasets
+accelerate
+
+# Load environment variables from .env file
+python-dotenv

src/cuda/huggingface/setup_environment.sh

@@ -0,0 +1,43 @@
+#!/bin/bash
+# Setup environment for running huggingface examples
+
+# Check if we are already in a virtual environment
+if [ -n "$VIRTUAL_ENV" ]; then
+    echo "Already in a virtual environment, skipping setup"
+    exit 0
+fi
+
+# Get the location of this script when sourcing
+if test -n "$BASH" ; then SCRIPT_LOC=$BASH_SOURCE
+elif test -n "$ZSH_NAME" ; then SCRIPT_LOC=${(%):-%x}
+else
+    echo "WARNING this script only tested with bash and zsh, use with caution with your shell at $SHELL"
+    if test -n "$TMOUT"; then SCRIPT_LOC=${.sh.file}
+    elif test ${0##*/} = dash; then x=$(lsof -p $$ -Fn0 | tail -1); SCRIPT_LOC=${x#n}
+    elif test -n "$FISH_VERSION" ; then SCRIPT_LOC=(status current-filename)
+    else echo "ERROR unknown shell, cannot determine script location" && return 1
+    fi
+fi
+
+# Get the directory of the script
+SCRIPT_DIR=${SCRIPT_LOC%/*}
+
+# Check if virtual environment exists
+if [ ! -d "$SCRIPT_DIR/.venv" ]; then
+    # Create virtual environment
+    python3 -m venv $SCRIPT_DIR/.venv
+    
+    # Activate virtual environment
+    source $SCRIPT_DIR/.venv/bin/activate
+    
+    # Install dependencies
+    pip install -r $SCRIPT_DIR/requirements.txt
+else
+    # Activate virtual environment
+    source $SCRIPT_DIR/.venv/bin/activate
+fi
+
+# Permission for python scripts
+chmod u+x $SCRIPT_DIR/*.py
+
+echo "Environment setup complete"

src/cuda/huggingface/util.py

@@ -0,0 +1,9 @@
+# Common flags for running huggingface examples
+import argparse
+
+def getCommonParser(is_parent=True):
+    # If is_parent is True, don't add the help flag to avoid conflict
+    parser = argparse.ArgumentParser(add_help=not is_parent)
+    parser.add_argument("--model_name", type=str, default="openai-community/gpt2-large")
+    parser.add_argument("--device", type=str, default="cuda")
+    return parser