Add HF Inference Endpoints custom handler

Auto-generated by Lovable so this LoRA adapter can be served as a custom-handler endpoint.

handler.py

@@ -0,0 +1,38 @@
+from typing import Any, Dict
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftConfig, PeftModel
+
+
+class EndpointHandler:
+    def __init__(self, path: str = ""):
+        cfg = PeftConfig.from_pretrained(path)
+        base = cfg.base_model_name_or_path
+        self.tokenizer = AutoTokenizer.from_pretrained(base)
+        if self.tokenizer.pad_token_id is None:
+            self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
+        model = AutoModelForCausalLM.from_pretrained(
+            base,
+            torch_dtype=torch.float16,
+            device_map="auto",
+        )
+        self.model = PeftModel.from_pretrained(model, path)
+        self.model.eval()
+
+    def __call__(self, data: Dict[str, Any]):
+        inputs = data.get("inputs", "")
+        params = data.get("parameters", {}) or {}
+        enc = self.tokenizer(inputs, return_tensors="pt").to(self.model.device)
+        with torch.no_grad():
+            out = self.model.generate(
+                **enc,
+                max_new_tokens=int(params.get("max_new_tokens", 256)),
+                temperature=float(params.get("temperature", 0.7)),
+                top_p=float(params.get("top_p", 0.9)),
+                do_sample=bool(params.get("do_sample", True)),
+                pad_token_id=self.tokenizer.pad_token_id,
+            )
+        text = self.tokenizer.decode(
+            out[0][enc["input_ids"].shape[1]:], skip_special_tokens=True
+        )
+        return [{"generated_text": text}]

requirements.txt

@@ -0,0 +1,6 @@
+transformers>=4.40.0
+peft>=0.10.0
+accelerate>=0.30.0
+safetensors>=0.4.0
+sentencepiece>=0.2.0
+bitsandbytes>=0.43.0