Create handler.py
Browse files- handler.py +56 -0
handler.py
ADDED
|
@@ -0,0 +1,56 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import Dict, Any
|
| 2 |
+
from transformers import JanusForConditionalGeneration, JanusProcessor
|
| 3 |
+
import torch, base64, io, PIL.Image as Image
|
| 4 |
+
|
| 5 |
+
class EndpointHandler:
|
| 6 |
+
"""
|
| 7 |
+
Works for:
|
| 8 |
+
• text → text chat completions
|
| 9 |
+
• text → image generation (pass {"generation_mode":"image"})
|
| 10 |
+
"""
|
| 11 |
+
def __init__(self, model_path: str):
|
| 12 |
+
self.processor = JanusProcessor.from_pretrained(
|
| 13 |
+
model_path, trust_remote_code=True
|
| 14 |
+
)
|
| 15 |
+
self.model = JanusForConditionalGeneration.from_pretrained(
|
| 16 |
+
model_path,
|
| 17 |
+
torch_dtype=torch.bfloat16, # fp16 also fine
|
| 18 |
+
device_map="auto",
|
| 19 |
+
load_in_4bit=True # comment out on bigger GPUs
|
| 20 |
+
)
|
| 21 |
+
|
| 22 |
+
# ---- each request lands here ----
|
| 23 |
+
def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
|
| 24 |
+
prompt = data.get("prompt") or data.get("inputs")
|
| 25 |
+
gen_mode = data.get("generation_mode", "text") # "text" | "image"
|
| 26 |
+
|
| 27 |
+
templ = self.processor.apply_chat_template(
|
| 28 |
+
[{"role": "user",
|
| 29 |
+
"content": [{"type": "text", "text": prompt}]}],
|
| 30 |
+
add_generation_prompt=True,
|
| 31 |
+
)
|
| 32 |
+
|
| 33 |
+
inputs = self.processor(
|
| 34 |
+
text=templ,
|
| 35 |
+
generation_mode=gen_mode,
|
| 36 |
+
return_tensors="pt"
|
| 37 |
+
).to(self.model.device)
|
| 38 |
+
|
| 39 |
+
out = self.model.generate(
|
| 40 |
+
**inputs,
|
| 41 |
+
generation_mode=gen_mode,
|
| 42 |
+
max_new_tokens=data.get("max_new_tokens", 128)
|
| 43 |
+
)
|
| 44 |
+
|
| 45 |
+
if gen_mode == "image":
|
| 46 |
+
img = self.processor.decode(out[0], output_type="pil")
|
| 47 |
+
return {"images": [self._pil_to_base64(img)]}
|
| 48 |
+
else:
|
| 49 |
+
return {"generated_text":
|
| 50 |
+
self.processor.decode(out[0], skip_special_tokens=True)}
|
| 51 |
+
|
| 52 |
+
@staticmethod
|
| 53 |
+
def _pil_to_base64(img: Image.Image) -> str:
|
| 54 |
+
buf = io.BytesIO()
|
| 55 |
+
img.save(buf, format="PNG")
|
| 56 |
+
return base64.b64encode(buf.getvalue()).decode()
|