Enable ZeroGPU on local backend, remove deprecated Docker files

HF requires at least one @spaces.GPU function to boot a gradio-sdk Space
on ZeroGPU hardware; decorate the local model's translate_batch and move
tensors to cuda when available. Drop the unused Dockerfile now that the
Space runs on the gradio SDK instead of docker.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Dockerfile

@@ -1,14 +0,0 @@
-FROM python:3.11-slim-bookworm
-
-ENV PYTHONUNBUFFERED=1
-
-WORKDIR /app
-
-COPY requirements.txt .
-RUN pip install --no-cache-dir -r requirements.txt
-
-COPY . .
-
-EXPOSE 7860
-
-CMD ["python", "app.py"]

engine/local_backend.py

@@ -11,8 +11,14 @@ version that ships AutoModelForSeq2SeqLM/AutoTokenizer.
 
 This backend ignores the editable translation prompt — it's a plain
 seq2seq model, not an instruction-following LLM.
+
+translate_batch is decorated with @spaces.GPU so this runs on HF ZeroGPU
+Spaces (which refuse to boot a gradio-sdk app with no @spaces.GPU function
+at all); the decorator is a no-op outside a ZeroGPU Space.
 """
 
+import spaces
+
 MODEL_ID = "billingsmoore/mlotsawa-ground-base"
 _PREFIX = "translate Tibetan to English: "
 
@@ -31,15 +37,18 @@ def _load():
     return _model, _tokenizer
 
 
+@spaces.GPU
 def translate_batch(texts: list[str]) -> list[str]:
     if not texts:
         return []
     import torch
     model, tokenizer = _load()
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    model.to(device)
     inputs = tokenizer(
         [_PREFIX + t for t in texts],
         return_tensors="pt", padding=True, truncation=True,
-    )
+    ).to(device)
     with torch.no_grad():
         outputs = model.generate(**inputs, max_length=300, num_beams=4, early_stopping=True)
     return [tokenizer.decode(o, skip_special_tokens=True) for o in outputs]

requirements.txt

@@ -1,4 +1,5 @@
 gradio>=6.9.0
+spaces
 google-genai
 python-dotenv
 transformers