deploy-endpoint.py

#!/usr/bin/env python3
"""Deploy SakThai Engine LoRA Adapter to HF Inference Endpoint

Run this AFTER the Kaggle notebook finishes training.
The adapter will be at: Nanthasit/sakthai-1.5b-lora-kaggle

Steps:
  1. Merge LoRA into base model
  2. Fix config.json (remove null quantization_config)
  3. Upload merged model to HF Hub
  4. Create Inference Endpoint
"""

import json, os, torch
from huggingface_hub import HfApi, login, hf_hub_download
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# ─── Config ─────────────────────────────────────────────
HF_TOKEN = os.environ["HF_TOKEN"]
BASE_MODEL = "Qwen/Qwen2.5-1.5B-Instruct"
ADAPTER_REPO = "Nanthasit/sakthai-1.5b-lora-kaggle"    # <- from Kaggle training
MERGED_REPO = "Nanthasit/sakthai-1.5b-merged-kaggle"    # <- upload target
ENDPOINT_NAME = "sakthai-1-5b-endpoint-v2"               # <- new endpoint

# ─── Auth ───────────────────────────────────────────────
login(token=HF_TOKEN, add_to_git_credential=True)
api = HfApi()

# ─── 1. Merge LoRA ──────────────────────────────────────
print("Loading base model...")
base = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)

print("Loading adapter...")
peft = PeftModel.from_pretrained(base, ADAPTER_REPO)
merged = peft.merge_and_unload()

print("Saving merged model...")
merged.save_pretrained("./merged-output")
tokenizer.save_pretrained("./merged-output")

# ─── 2. Fix config.json ────────────────────────────────
config_path = "./merged-output/config.json"
config = json.loads(open(config_path).read())
config.pop("quantization_config", None)  # <- required fix!
config["torch_dtype"] = "bfloat16"
with open(config_path, "w") as f:
    json.dump(config, f, indent=2)
print("config.json fixed")

# ─── 3. Upload to Hub ───────────────────────────────────
api.create_repo(repo_id=MERGED_REPO, repo_type="model", exist_ok=True)
api.upload_folder(
    folder_path="./merged-output",
    repo_id=MERGED_REPO,
    commit_message="Merged model from Kaggle LoRA training",
)
print(f"Merged model: https://huggingface.co/{MERGED_REPO}")

# ─── 4. Create Inference Endpoint ───────────────────────
endpoint = api.create_inference_endpoint(
    name=ENDPOINT_NAME,
    repository=MERGED_REPO,
    accelerator="gpu",
    instance_type="nvidia-a10g",
    instance_size="x1",
    region="us-east-1",
    vendor="aws",
    framework="pytorch",
    task="text-generation",
    min_replica=0,                  # scale-to-zero
    max_replica=1,
    scale_to_zero_timeout=15,
)
print(f"Endpoint: {endpoint.url}")
print("✅ Done! Ready for inference.")