Files
model_garden/community-content/vertex_model_garden/model_oss/peft/handler.py
T

251 lines
8.8 KiB
Python

"""Custom handler for huggingface/peft models."""
# pylint: disable=g-importing-member
# pylint: disable=logging-fstring-interpolation
import logging
import os
from typing import Any, List
from absl import logging
from diffusers import DPMSolverMultistepScheduler
from diffusers import StableDiffusionPipeline
from peft import PeftModel
from PIL import Image
import torch
import transformers
from transformers import AutoModelForCausalLM
from transformers import AutoModelForSequenceClassification
from transformers import AutoTokenizer
from transformers import BitsAndBytesConfig
from ts.torch_handler.base_handler import BaseHandler
from util import constants
from util import fileutils
from util import image_format_converter
# Tasks
TEXT_TO_IMAGE_LORA = "text-to-image-lora"
SEQUENCE_CLASSIFICATION_LORA = "sequence-classification-lora"
CAUSAL_LANGUAGE_MODELING_LORA = "causal-language-modeling-lora"
INSTRUCT_LORA = "instruct-lora"
# Inference parameters.
_NUM_INFERENCE_STEPS = 25
_MAX_LENGTH_DEFAULT = 200
_TOP_K_DEFAULT = 10
class PeftHandler(BaseHandler):
"""Custom handler for Peft models."""
def initialize(self, context: Any):
"""Initializes the handler."""
logging.info("Start to initialize the PEFT handler.")
properties = context.system_properties
self.map_location = (
"cuda"
if torch.cuda.is_available() and properties.get("gpu_id") is not None
else "cpu"
)
self.device = torch.device(
self.map_location + ":" + str(properties.get("gpu_id"))
if torch.cuda.is_available() and properties.get("gpu_id") is not None
else self.map_location
)
self.manifest = context.manifest
self.precision_mode = os.environ.get(
"PRECISION_LOADING_MODE", constants.PRECISION_MODE_16
)
self.task = os.environ.get("TASK", CAUSAL_LANGUAGE_MODELING_LORA)
self.base_model_id = os.environ.get(
"BASE_MODEL_ID", "openlm-research/open_llama_7b"
)
if fileutils.is_gcs_path(self.base_model_id):
fileutils.download_gcs_dir_to_local(
self.base_model_id,
constants.LOCAL_BASE_MODEL_DIR,
skip_hf_model_bin=True,
)
self.base_model_id = constants.LOCAL_BASE_MODEL_DIR
self.finetuned_lora_model_path = os.environ.get(
"FINETUNED_LORA_MODEL_PATH", ""
)
if fileutils.is_gcs_path(self.finetuned_lora_model_path):
fileutils.download_gcs_dir_to_local(
self.finetuned_lora_model_path, constants.LOCAL_MODEL_DIR
)
self.finetuned_lora_model_path = constants.LOCAL_MODEL_DIR
logging.info(
f"Using task:{self.task}, base model:{self.base_model_id}, lora model:"
f" {self.finetuned_lora_model_path}, and precision"
f" {self.precision_mode}."
)
self.pipeline = None
self.model = None
self.tokenizer = None
if self.task == TEXT_TO_IMAGE_LORA:
pipeline = StableDiffusionPipeline.from_pretrained(
self.base_model_id, torch_dtype=torch.float16
)
logging.debug("Initialized the base model for text to image.")
pipeline.scheduler = DPMSolverMultistepScheduler.from_config(
pipeline.scheduler.config
)
logging.debug("Initialized the scheduler for text to image.")
if self.finetuned_lora_model_path:
pipeline.unet.load_attn_procs(self.finetuned_lora_model_path)
logging.debug("Initialized the LoRA model for text to image.")
# This is to reduce GPU memory requirements.
pipeline.enable_xformers_memory_efficient_attention()
pipeline = pipeline.to(self.map_location)
# Reduces memory footprint.
pipeline.enable_attention_slicing()
self.pipeline = pipeline
logging.info("Initialized the text to image pipelines.")
elif self.task == SEQUENCE_CLASSIFICATION_LORA:
tokenizer = AutoTokenizer.from_pretrained(self.base_model_id)
logging.debug("Initialized the tokenizer for sequence classification.")
model = AutoModelForSequenceClassification.from_pretrained(
self.base_model_id, torch_dtype=torch.float16
)
logging.debug("Initialized the base model for sequence classification.")
if self.finetuned_lora_model_path:
model = PeftModel.from_pretrained(model, self.finetuned_lora_model_path)
logging.debug("Initialized the LoRA model for sequence classification.")
model.to(self.map_location)
self.model = model
self.tokenizer = tokenizer
elif (
self.task == CAUSAL_LANGUAGE_MODELING_LORA or self.task == INSTRUCT_LORA
):
tokenizer = AutoTokenizer.from_pretrained(self.base_model_id)
logging.debug("Initialized the tokenizer.")
if self.task == CAUSAL_LANGUAGE_MODELING_LORA:
if self.precision_mode == constants.PRECISION_MODE_32:
model = AutoModelForCausalLM.from_pretrained(
self.base_model_id,
return_dict=True,
torch_dtype=torch.float32,
device_map="auto",
)
elif self.precision_mode == constants.PRECISION_MODE_16:
model = AutoModelForCausalLM.from_pretrained(
self.base_model_id,
return_dict=True,
torch_dtype=torch.bfloat16,
device_map="auto",
)
elif self.precision_mode == constants.PRECISION_MODE_8:
quantization_config = BitsAndBytesConfig(
load_in_8bit=True, int8_threshold=0
)
model = AutoModelForCausalLM.from_pretrained(
self.base_model_id,
return_dict=True,
torch_dtype=torch.float16,
device_map="auto",
quantization_config=quantization_config,
)
else:
quantization_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.bfloat16,
)
model = AutoModelForCausalLM.from_pretrained(
self.base_model_id,
return_dict=True,
device_map="auto",
torch_dtype=torch.bfloat16,
quantization_config=quantization_config,
)
else:
model = AutoModelForCausalLM.from_pretrained(
self.base_model_id,
return_dict=True,
torch_dtype=torch.bfloat16,
trust_remote_code=True,
device_map="auto",
)
logging.debug("Initialized the base model.")
if self.finetuned_lora_model_path:
model = PeftModel.from_pretrained(model, self.finetuned_lora_model_path)
logging.debug("Initialized the LoRA model.")
pipeline = transformers.pipeline(
"text-generation",
model=model,
tokenizer=tokenizer,
)
self.tokenizer = tokenizer
self.pipeline = pipeline
else:
raise ValueError(f"Invalid TASK: {self.task}")
self.initialized = True
logging.info("The PEFT handler was initialized.")
def preprocess(self, data: Any) -> Any:
"""Preprocesses input data."""
# Assumes that the parameters are same in one request. We parse the
# parameters from the first instance for all instances in one request.
max_length = _MAX_LENGTH_DEFAULT
top_k = _TOP_K_DEFAULT
prompts = [item["prompt"] for item in data]
if "max_length" in data[0]:
max_length = data[0]["max_length"]
if "top_k" in data[0]:
top_k = data[0]["top_k"]
return prompts, max_length, top_k
def inference(self, data: Any, *args, **kwargs) -> List[Image.Image]:
"""Runs the inference."""
prompts, max_length, top_k = data
logging.debug(
f"Inference prompts={prompts}, max_length={max_length}, top_k={top_k}."
)
if self.task == TEXT_TO_IMAGE_LORA:
predicted_results = self.pipeline(
prompt=prompts, num_inference_steps=_NUM_INFERENCE_STEPS
).images
elif self.task == SEQUENCE_CLASSIFICATION_LORA:
encoded_input = self.tokenizer(prompts, return_tensors="pt")
encoded_input.to(self.map_location)
with torch.no_grad():
outputs = self.model(**encoded_input)
predictions = outputs.logits.argmax(dim=-1)
predicted_results = predictions.tolist()
elif (
self.task == CAUSAL_LANGUAGE_MODELING_LORA or self.task == INSTRUCT_LORA
):
predicted_results = self.pipeline(
prompts,
max_length=max_length,
do_sample=True,
top_k=top_k,
num_return_sequences=1,
eos_token_id=self.tokenizer.eos_token_id,
)
else:
raise ValueError(f"Invalid TASK: {self.task}")
return predicted_results
def postprocess(self, data: Any) -> List[str]:
"""Postprocesses output data."""
if self.task == TEXT_TO_IMAGE_LORA:
# Converts the images to base64 string.
outputs = [
image_format_converter.image_to_base64(image) for image in data
]
else:
outputs = data
return outputs
# pylint: enable=logging-fstring-interpolation