141 lines
6.5 KiB
Python
141 lines
6.5 KiB
Python
import modal
|
|
from modal import App, Volume, Image
|
|
|
|
import logging
|
|
logging.basicConfig(level=logging.INFO)
|
|
|
|
# ─────────────────────────────────────────────────────────────────────────────
|
|
# Constants
|
|
# ─────────────────────────────────────────────────────────────────────────────
|
|
|
|
GPU = "T4" # Use a T4 GPU for inference
|
|
CACHE_PATH = "/cache" # Mount point for the Modal volume
|
|
|
|
# Hugging Face model references
|
|
BASE_MODEL = "meta-llama/Meta-Llama-3.1-8B"
|
|
FINETUNED_MODEL = "ed-donner/pricer-2024-09-13_13.04.39"
|
|
REVISION = "e8d637df551603dc86cd7a1598a8f44af4d7ae36" # Commit of the fine-tuned model
|
|
|
|
# Local cache paths (inside the volume)
|
|
BASE_MODEL_DIR = f"{CACHE_PATH}/llama_base_model"
|
|
FINETUNED_MODEL_DIR = f"{CACHE_PATH}/llama_finetuned_model"
|
|
|
|
# ─────────────────────────────────────────────────────────────────────────────
|
|
# Structure
|
|
# ─────────────────────────────────────────────────────────────────────────────
|
|
|
|
# Container (App: llm-ft-pricer)
|
|
# ├── /app ← Code + installed Python packages (from image)
|
|
# ├── /cache ← Mounted Modal volume (`hf-hub-cache`)
|
|
# │ └── meta-llama/Meta-Llama-3.1-8B/... ← HuggingFace model files downloaded via snapshot_download
|
|
|
|
|
|
|
|
QUESTION = "How much does this cost to the nearest dollar?"
|
|
PREFIX = "Price is $" # Used to parse generated output
|
|
|
|
# ─────────────────────────────────────────────────────────────────────────────
|
|
# Modal App, Image, Volume, Secrets
|
|
# ─────────────────────────────────────────────────────────────────────────────
|
|
|
|
app = modal.App("llm-ft-pricer") # Define the Modal app
|
|
|
|
image = (
|
|
Image.debian_slim()
|
|
.pip_install("huggingface", "torch", "transformers", "bitsandbytes", "accelerate", "peft") # All needed libraries
|
|
.env({"HF_HUB_CACHE": CACHE_PATH}) # Hugging Face will store model files in /cache
|
|
)
|
|
|
|
cache_vol = modal.Volume.from_name("hf-hub-cache", create_if_missing=True) # Persisted volume for caching models
|
|
secrets = [modal.Secret.from_name("HF_TOKEN")] # Hugging Face auth token
|
|
|
|
# ─────────────────────────────────────────────────────────────────────────────
|
|
# Modal Class: Pricer
|
|
# ─────────────────────────────────────────────────────────────────────────────
|
|
|
|
# All methods in this class run inside the container with the image, volume, secrets, and GPU you configured.
|
|
@app.cls(
|
|
image=image,
|
|
secrets=secrets,
|
|
volumes={CACHE_PATH: cache_vol}, # Mount volume into /cache
|
|
gpu=GPU,
|
|
timeout=1800, # 30-minute max runtime
|
|
min_containers=0, # = 1 : Keeping one container warm uses credits continuously if you forget to stop it.
|
|
scaledown_window=300, # Shuts down the container
|
|
)
|
|
class Pricer:
|
|
@modal.enter()
|
|
def setup(self):
|
|
import os, torch
|
|
import logging
|
|
from huggingface_hub import snapshot_download
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
|
|
from peft import PeftModel
|
|
|
|
# Create cache path if it doesn't exist
|
|
os.makedirs(CACHE_PATH, exist_ok=True)
|
|
|
|
# Download base and fine-tuned models into volume
|
|
logging.info("Downloading base model...")
|
|
snapshot_download(BASE_MODEL, local_dir=BASE_MODEL_DIR)
|
|
|
|
logging.info("Downloading fine-tuned model...")
|
|
snapshot_download(FINETUNED_MODEL, revision=REVISION, local_dir=FINETUNED_MODEL_DIR)
|
|
|
|
# Quantization config (4-bit)
|
|
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 tokenizer
|
|
self.tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_DIR)
|
|
self.tokenizer.pad_token = self.tokenizer.eos_token
|
|
self.tokenizer.padding_side = "right"
|
|
|
|
# Load base model (quantized)
|
|
base_model = AutoModelForCausalLM.from_pretrained(
|
|
BASE_MODEL_DIR,
|
|
quantization_config=quant_config,
|
|
device_map="auto"
|
|
)
|
|
|
|
# Apply fine-tuned weights
|
|
self.fine_tuned_model = PeftModel.from_pretrained(
|
|
base_model,
|
|
FINETUNED_MODEL_DIR,
|
|
revision=REVISION
|
|
)
|
|
self.fine_tuned_model.generation_config.pad_token_id = self.tokenizer.pad_token_id
|
|
|
|
@modal.method()
|
|
def price(self, description: str) -> float:
|
|
import re, torch
|
|
from transformers import set_seed
|
|
|
|
set_seed(42) # Deterministic output
|
|
|
|
# Construct prompt
|
|
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")
|
|
|
|
# Generate model output (max 5 tokens)
|
|
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])
|
|
|
|
# Extract number after "Price is $"
|
|
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 # Return parsed price or 0 if not found
|
|
|
|
|