74 lines
2.7 KiB
Python
74 lines
2.7 KiB
Python
|
"""Custom endpoint handler for HF Inference Endpoints.
|
||
|
|
|
||
|
|
Bypasses the default huggingface_inference_toolkit auto-pipeline (which was
|
||
|
|
crashing at model-load on this Qwen2.5-Coder merge). Loads explicitly in bf16,
|
||
|
|
handles both plain-string and chat-message inputs, returns the generated text
|
||
|
|
in the shape the MCP client expects: [{"generated_text": "..."}].
|
||
|
|
"""
|
||
|
|
from __future__ import annotations
|
||
|
|
|
||
|
|
from typing import Any
|
||
|
|
|
||
|
|
import torch
|
||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
|
|
||
|
|
|
||
|
|
class EndpointHandler:
|
||
|
|
def __init__(self, path: str = ""):
|
||
|
|
print(f"[handler] Loading tokenizer from {path}")
|
||
|
|
self.tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True)
|
||
|
|
if self.tokenizer.pad_token is None:
|
||
|
|
self.tokenizer.pad_token = self.tokenizer.eos_token
|
||
|
|
|
||
|
|
print(f"[handler] Loading model from {path} in bfloat16")
|
||
|
|
self.model = AutoModelForCausalLM.from_pretrained(
|
||
|
|
path,
|
||
|
|
torch_dtype=torch.bfloat16,
|
||
|
|
device_map="auto",
|
||
|
|
trust_remote_code=True,
|
||
|
|
)
|
||
|
|
self.model.eval()
|
||
|
|
print("[handler] Ready")
|
||
|
|
|
||
|
|
def __call__(self, data: dict[str, Any]) -> list[dict[str, str]]:
|
||
|
|
inputs = data.get("inputs", "")
|
||
|
|
params = data.get("parameters", {}) or {}
|
||
|
|
|
||
|
|
max_new_tokens = int(params.get("max_new_tokens", 1500))
|
||
|
|
temperature = float(params.get("temperature", 0.1))
|
||
|
|
top_p = float(params.get("top_p", 0.95))
|
||
|
|
do_sample = bool(params.get("do_sample", temperature > 0.01))
|
||
|
|
return_full_text = bool(params.get("return_full_text", False))
|
||
|
|
|
||
|
|
# Accept either a plain string or a list of chat messages.
|
||
|
|
if isinstance(inputs, list) and inputs and isinstance(inputs[0], dict):
|
||
|
|
prompt_text = self.tokenizer.apply_chat_template(
|
||
|
|
inputs, tokenize=False, add_generation_prompt=True
|
||
|
|
)
|
||
|
|
else:
|
||
|
|
prompt_text = inputs
|
||
|
|
|
||
|
|
tokenized = self.tokenizer(
|
||
|
|
prompt_text,
|
||
|
|
return_tensors="pt",
|
||
|
|
truncation=True,
|
||
|
|
max_length=8192,
|
||
|
|
).to(self.model.device)
|
||
|
|
|
||
|
|
with torch.no_grad():
|
||
|
|
outputs = self.model.generate(
|
||
|
|
**tokenized,
|
||
|
|
max_new_tokens=max_new_tokens,
|
||
|
|
temperature=temperature if do_sample else 1.0,
|
||
|
|
top_p=top_p,
|
||
|
|
do_sample=do_sample,
|
||
|
|
pad_token_id=self.tokenizer.eos_token_id,
|
||
|
|
eos_token_id=self.tokenizer.eos_token_id,
|
||
|
|
)
|
||
|
|
|
||
|
|
prompt_len = tokenized["input_ids"].shape[-1]
|
||
|
|
new_tokens = outputs[0] if return_full_text else outputs[0][prompt_len:]
|
||
|
|
text = self.tokenizer.decode(new_tokens, skip_special_tokens=True)
|
||
|
|
|
||
|
|
return [{"generated_text": text}]
|