You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 

93 lines
3.4 KiB

import modal
from pathlib import PurePosixPath
# Setup - define our infrastructure with code!
app = modal.App("pricer-service")
image = modal.Image.debian_slim().pip_install(
"huggingface", "torch", "transformers", "bitsandbytes",
"accelerate", "peft", "huggingface_hub[hf_transfer]"
).env({"HF_HUB_ENABLE_HF_TRANSFER": "1"})
hf_cache_vol = modal.Volume.from_name("hf-cache", create_if_missing=True)
secrets = [modal.Secret.from_name("huggingface-secret")]
# Constants
GPU = "T4"
BASE_MODEL = "meta-llama/Meta-Llama-3.1-8B"
PROJECT_NAME = "pricer"
HF_USER = "ed-donner" # your HF name here! Or use mine if you just want to reproduce my results.
RUN_NAME = "2024-09-13_13.04.39"
PROJECT_RUN_NAME = f"{PROJECT_NAME}-{RUN_NAME}"
REVISION = "e8d637df551603dc86cd7a1598a8f44af4d7ae36"
FINETUNED_MODEL = f"{HF_USER}/{PROJECT_RUN_NAME}"
# Define cache locations - this will map to the volume created
MODEL_DIR = PurePosixPath("/models")
BASE_DIR = MODEL_DIR / BASE_MODEL
FINETUNED_DIR = MODEL_DIR / FINETUNED_MODEL
QUESTION = "How much does this cost to the nearest dollar?"
PREFIX = "Price is $"
@app.cls(image=image, secrets=secrets, gpu=GPU, timeout=1800, volumes={MODEL_DIR: hf_cache_vol})
class Pricer:
@modal.enter()
def setup(self):
import torch
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel
# Download and cache model files to the volume
snapshot_download(BASE_MODEL, local_dir=BASE_DIR)
snapshot_download(FINETUNED_MODEL, revision=REVISION, local_dir=FINETUNED_DIR)
# Quant Config
quant_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_use_double_quant=True,
bnb_4bit_compute_dtype=torch.bfloat16,
bnb_4bit_quant_type="nf4"
)
# Load model and tokenizer
self.tokenizer = AutoTokenizer.from_pretrained(BASE_DIR)
self.tokenizer.pad_token = self.tokenizer.eos_token
self.tokenizer.padding_side = "right"
self.base_model = AutoModelForCausalLM.from_pretrained(
BASE_DIR,
quantization_config=quant_config,
device_map="auto"
)
self.fine_tuned_model = PeftModel.from_pretrained(self.base_model, FINETUNED_DIR, revision=REVISION)
@modal.method()
def price(self, description: str) -> float:
import re, torch
from transformers import set_seed
set_seed(42)
prompt = f"{QUESTION}\n\n{description}\n\n{PREFIX}"
inputs = self.tokenizer.encode(prompt, return_tensors="pt").to("cuda")
attention_mask = torch.ones(inputs.shape, device="cuda")
outputs = self.fine_tuned_model.generate(inputs, attention_mask=attention_mask, max_new_tokens=5, num_return_sequences=1)
result = self.tokenizer.decode(outputs[0])
contents = result.split("Price is $")[1]
contents = contents.replace(',','')
match = re.search(r"[-+]?\d*\.\d+|\d+", contents)
return float(match.group()) if match else 0
@modal.method()
def wake_up(self) -> str:
return "ok"
## Keep Pricer warm so it's faster to respond to requests
@app.function(schedule=modal.Period(seconds=50))
def pricer_wake_up() -> str:
Pricer = modal.Cls.from_name("pricer-service", "Pricer")
return Pricer().wake_up.remote()