Remove internal serving handler from public release

handler.py

@@ -1,99 +0,0 @@
-import os
-import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer
-
-# Keep only the last round (1 user+assistant pair) + the current user message
-MAX_HISTORY_MESSAGES = 3
-
-
-class EndpointHandler:
-    """
-    Hugging Face Inference Endpoints custom handler.
-
-    Expects input like:
-      - {"inputs": "hello"}                                  -> single-turn, auto-wrapped with chat template
-      - {"inputs": [{"role":"user","content":"hello"}, ...]}  -> multi-turn chat (messages list)
-      - {"inputs": "hello", "parameters": {"raw": true}}     -> sent as-is (no template)
-
-    Optional:
-      - {"parameters": {"max_new_tokens": 512, "temperature": 0.7, ...}}
-    """
-
-    def __init__(self, path: str = ""):
-        model_dir = path or os.getenv("HF_MODEL_DIR", ".")
-        self.device = "cuda" if torch.cuda.is_available() else "cpu"
-
-        self.tokenizer = AutoTokenizer.from_pretrained(model_dir, trust_remote_code=True)
-
-        # Ensure pad token exists (common for causal LMs)
-        if self.tokenizer.pad_token is None:
-            self.tokenizer.pad_token = self.tokenizer.eos_token
-            self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
-
-
-        self.model = AutoModelForCausalLM.from_pretrained(
-            model_dir,
-            torch_dtype="auto",
-            device_map="auto" if torch.cuda.is_available() else None,
-            trust_remote_code=True,
-        )
-        self.model.eval()
-
-    def __call__(self, data: dict) -> dict:
-        inputs = data.get("inputs", data)
-        params = data.get("parameters", {}) or {}
-
-        raw = bool(params.pop("raw", False))
-
-        if raw:
-            if not isinstance(inputs, str):
-                raise ValueError("raw mode requires inputs to be a string.")
-            prompt = inputs
-        elif isinstance(inputs, list):
-            # Multi-turn: inputs is a list of {"role": ..., "content": ...}
-            inputs = inputs[-MAX_HISTORY_MESSAGES:]
-            prompt = self.tokenizer.apply_chat_template(
-                inputs, tokenize=False,
-            )
-        elif isinstance(inputs, str):
-            # Single-turn: wrap in a one-message list
-            prompt = self.tokenizer.apply_chat_template(
-                [{"role": "user", "content": inputs}],
-                tokenize=False,
-            )
-        else:
-            raise ValueError("inputs must be a string or a list of messages.")
-
-        enc = self.tokenizer(
-            prompt,
-            return_tensors="pt",
-            padding=False,
-            truncation=True,
-        )
-        input_ids = enc["input_ids"].to(self.model.device)
-        attention_mask = enc.get("attention_mask", torch.ones_like(input_ids)).to(self.model.device)
-
-        gen_kwargs = {
-            "max_new_tokens": min(int(params.pop("max_new_tokens", 512)), 512),
-            "do_sample": bool(params.pop("do_sample", True)),
-            "temperature": float(params.pop("temperature", 0.7)),
-            "top_p": float(params.pop("top_p", 0.95)),
-            "repetition_penalty": float(params.pop("repetition_penalty", 1.2)),
-            "eos_token_id": self.tokenizer.eos_token_id,
-            "pad_token_id": self.tokenizer.pad_token_id,
-        }
-        gen_kwargs.update(params)
-
-        with torch.no_grad():
-            out = self.model.generate(
-                input_ids=input_ids,
-                attention_mask=attention_mask,
-                **gen_kwargs,
-            )
-
-        # Return only newly generated tokens (your current behavior)
-        new_tokens = out[0, input_ids.shape[-1]:]
-        text = self.tokenizer.decode(new_tokens, skip_special_tokens=True)
-
-        return {"generated_text": text}
-