handler.py

import contextlib, io, base64, torch
from PIL import Image
import open_clip
from reparam import reparameterize_model

class EndpointHandler:
    def __init__(self, path: str = ""):
        self.device = "cuda" if torch.cuda.is_available() else "cpu"

        # Fix 1: Load weights directly from the web, just like local script
        # This guarantees the weights are identical.
        model, _, self.preprocess = open_clip.create_model_and_transforms(
            "MobileCLIP-B", pretrained='datacompdr'
        )
        model.eval()
        self.model = reparameterize_model(model)  # fuse branches

        self.tokenizer = open_clip.get_tokenizer("MobileCLIP-B")
        self.model.to(self.device)

        # Fix 2: Explicitly set model to half-precision if on CUDA
        # This matches the behavior of torch.set_default_dtype(torch.float16)
        if self.device == "cuda":
            self.model.to(torch.float16)

    def __call__(self, data):
        payload = data.get("inputs", data)
        img_b64 = payload["image"]
        labels  = payload.get("candidate_labels", [])
        if not labels:
            return {"error": "candidate_labels list is empty"}

        # ---------------- decode inputs ----------------
        image = Image.open(io.BytesIO(base64.b64decode(img_b64))).convert("RGB")
        img_tensor = self.preprocess(image).unsqueeze(0).to(self.device)

        # The preprocessor might output float32, so ensure tensor matches model dtype
        if self.device == "cuda":
            img_tensor = img_tensor.to(torch.float16)

        text_tokens = self.tokenizer(labels).to(self.device)

        # ---------------- forward pass -----------------
        # No need for autocast if everything is already float16
        with torch.no_grad():
            img_feat = self.model.encode_image(img_tensor)
            txt_feat = self.model.encode_text(text_tokens)
            img_feat /= img_feat.norm(dim=-1, keepdim=True)
            txt_feat /= txt_feat.norm(dim=-1, keepdim=True)
            probs = (100 * img_feat @ txt_feat.T).softmax(dim=-1)[0].cpu().tolist()

        return [
            {"label": l, "score": float(p)}
            for l, p in sorted(zip(labels, probs), key=lambda x: x[1], reverse=True)
        ]
# import contextlib, io, base64, torch
# from PIL import Image
# import open_clip
# from reparam import reparameterize_model


# class EndpointHandler:
#     def __init__(self, path: str = ""):
#         # You can also pass pretrained='datacompdr' to let OpenCLIP download
#         weights = f"{path}/mobileclip_b.pt"
#         self.model, _, self.preprocess = open_clip.create_model_and_transforms(
#             "MobileCLIP-B", pretrained=weights
#         )
#         self.model.eval()
#         self.model = reparameterize_model(self.model)   # *** fuse branches ***

#         self.device = "cuda" if torch.cuda.is_available() else "cpu"
#         self.model.to(self.device)
#         self.tokenizer = open_clip.get_tokenizer("MobileCLIP-B")

#     def __call__(self, data):
#         payload = data.get("inputs", data)
#         img_b64 = payload["image"]
#         labels  = payload.get("candidate_labels", [])
#         if not labels:
#             return {"error": "candidate_labels list is empty"}

#         # ---------------- decode inputs ----------------
#         image = Image.open(io.BytesIO(base64.b64decode(img_b64))).convert("RGB")
#         img_tensor  = self.preprocess(image).unsqueeze(0).to(self.device)
#         text_tokens = self.tokenizer(labels).to(self.device)

#         # ---------------- forward pass -----------------
#         autocast_ctx = (
#             torch.cuda.amp.autocast if self.device.startswith("cuda") else contextlib.nullcontext
#         )
#         with torch.no_grad(), autocast_ctx():
#             img_feat = self.model.encode_image(img_tensor)
#             txt_feat = self.model.encode_text(text_tokens)
#             img_feat /= img_feat.norm(dim=-1, keepdim=True)
#             txt_feat /= txt_feat.norm(dim=-1, keepdim=True)
#             probs = (100 * img_feat @ txt_feat.T).softmax(dim=-1)[0].tolist()

#         return [
#             {"label": l, "score": float(p)}
#             for l, p in sorted(zip(labels, probs), key=lambda x: x[1], reverse=True)
#         ]