Browse Source

Removed modal.build() in favor of using modal.Volume to cache model files. Added warming code as an app.function()

pull/261/head
Shay Harding 2 months ago
parent
commit
ae0e2914cd
  1. 54
      week8/pricer_service2.py

54
week8/pricer_service2.py

@ -1,14 +1,18 @@
import modal import modal
from modal import App, Volume, Image from pathlib import PurePosixPath
# Setup - define our infrastructure with code! # Setup - define our infrastructure with code!
app = modal.App("pricer-service") app = modal.App("pricer-service")
image = Image.debian_slim().pip_install("huggingface", "torch", "transformers", "bitsandbytes", "accelerate", "peft")
secrets = [modal.Secret.from_name("hf-secret")]
# Constants 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" GPU = "T4"
BASE_MODEL = "meta-llama/Meta-Llama-3.1-8B" BASE_MODEL = "meta-llama/Meta-Llama-3.1-8B"
PROJECT_NAME = "pricer" PROJECT_NAME = "pricer"
@ -17,30 +21,28 @@ RUN_NAME = "2024-09-13_13.04.39"
PROJECT_RUN_NAME = f"{PROJECT_NAME}-{RUN_NAME}" PROJECT_RUN_NAME = f"{PROJECT_NAME}-{RUN_NAME}"
REVISION = "e8d637df551603dc86cd7a1598a8f44af4d7ae36" REVISION = "e8d637df551603dc86cd7a1598a8f44af4d7ae36"
FINETUNED_MODEL = f"{HF_USER}/{PROJECT_RUN_NAME}" FINETUNED_MODEL = f"{HF_USER}/{PROJECT_RUN_NAME}"
MODEL_DIR = "hf-cache/"
BASE_DIR = MODEL_DIR + BASE_MODEL # Define cache locations - this will map to the volume created
FINETUNED_DIR = MODEL_DIR + FINETUNED_MODEL 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?" QUESTION = "How much does this cost to the nearest dollar?"
PREFIX = "Price is $" PREFIX = "Price is $"
@app.cls(image=image, secrets=secrets, gpu=GPU, timeout=1800) @app.cls(image=image, secrets=secrets, gpu=GPU, timeout=1800, volumes={MODEL_DIR: hf_cache_vol})
class Pricer: class Pricer:
@modal.build()
def download_model_to_folder(self):
from huggingface_hub import snapshot_download
import os
os.makedirs(MODEL_DIR, exist_ok=True)
snapshot_download(BASE_MODEL, local_dir=BASE_DIR)
snapshot_download(FINETUNED_MODEL, revision=REVISION, local_dir=FINETUNED_DIR)
@modal.enter() @modal.enter()
def setup(self): def setup(self):
import os
import torch import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig, set_seed from huggingface_hub import snapshot_download
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel 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
quant_config = BitsAndBytesConfig( quant_config = BitsAndBytesConfig(
load_in_4bit=True, load_in_4bit=True,
@ -50,7 +52,6 @@ class Pricer:
) )
# Load model and tokenizer # Load model and tokenizer
self.tokenizer = AutoTokenizer.from_pretrained(BASE_DIR) self.tokenizer = AutoTokenizer.from_pretrained(BASE_DIR)
self.tokenizer.pad_token = self.tokenizer.eos_token self.tokenizer.pad_token = self.tokenizer.eos_token
self.tokenizer.padding_side = "right" self.tokenizer.padding_side = "right"
@ -65,11 +66,8 @@ class Pricer:
@modal.method() @modal.method()
def price(self, description: str) -> float: def price(self, description: str) -> float:
import os import re, torch
import re from transformers import set_seed
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig, set_seed
from peft import PeftModel
set_seed(42) set_seed(42)
prompt = f"{QUESTION}\n\n{description}\n\n{PREFIX}" prompt = f"{QUESTION}\n\n{description}\n\n{PREFIX}"
@ -87,3 +85,9 @@ class Pricer:
def wake_up(self) -> str: def wake_up(self) -> str:
return "ok" 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()

Loading…
Cancel
Save