Create app.py

app.py

@@ -0,0 +1,55 @@
+import torch
+from transformers import AutoTokenizer, AutoModel
+from fastapi import FastAPI, HTTPException
+from pydantic import BaseModel
+import os
+
+MODEL_NAME = os.getenv("MODEL_NAME", "jhu-clsp/mmBERT-base")
+
+app = FastAPI(title="ModernBERT Embedding API", version="1.0.0")
+
+print("Loading model:", MODEL_NAME)
+
+tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
+model = AutoModel.from_pretrained(MODEL_NAME)
+model.eval()
+
+
+class EmbedRequest(BaseModel):
+    text: str
+
+
+@app.get("/health")
+def health():
+    return {"status": "ok", "model": MODEL_NAME}
+
+
+@app.post("/embed")
+def embed(req: EmbedRequest):
+    text = (req.text or "").strip()
+
+    if not text:
+        raise HTTPException(status_code=400, detail="Empty text")
+
+    with torch.no_grad():
+        inputs = tokenizer(
+            text,
+            padding=True,
+            truncation=True,
+            max_length=512,
+            return_tensors="pt",
+        )
+
+        outputs = model(**inputs)
+
+        mask = inputs["attention_mask"].unsqueeze(-1)
+        embeddings = (outputs.last_hidden_state * mask).sum(dim=1) / mask.sum(dim=1)
+
+        emb = embeddings[0].tolist()
+
+    return {
+        "model": MODEL_NAME,
+        "dim": len(emb),
+        "preview_first_8": [round(x, 4) for x in emb[:8]],
+        "embedding": emb,
+    }