Current changes

.gitignore

@@ -1,2 +1,3 @@
 translator_app/.venv
-backend
+translator_app
+translator_app/db.sqlite3

Dockerfile

@@ -33,4 +33,6 @@ RUN apt-get update && apt-get install -y --no-install-recommends patchelf \
 COPY backend/ ./
 COPY --from=model /m/models ./models
 EXPOSE 7860
-CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "7860"]
+COPY entrypoint.sh .
+RUN chmod +x entrypoint.sh
+CMD ["./entrypoint.sh"]

backend/main.py

@@ -7,6 +7,7 @@ import os
 import httpx
 from fastapi import FastAPI, Header, HTTPException
 from fastapi.middleware.cors import CORSMiddleware
+from scalar_fastapi import get_scalar_api_reference
 from pydantic import BaseModel, Field
 
 import providers
@@ -33,7 +34,9 @@ class TranslateRequest(BaseModel):
     text: str = Field(..., description="Text to translate.")
     source: str = Field(..., description="Source FLORES-200 code, e.g. 'eng_Latn'.")
     target: str = Field(..., description="Target FLORES-200 code, e.g. 'kor_Hang'.")
-    engine: str | None = Field(None, description="Engine id; defaults to first available.")
+    engine: str | None = Field(
+        None, description="Engine id; defaults to first available."
+    )
 
 
 class TranslateResponse(BaseModel):
@@ -69,7 +72,10 @@ def translate_endpoint(
 ) -> TranslateResponse:
     if not req.text.strip():
         return TranslateResponse(
-            translation="", source=req.source, target=req.target, engine=req.engine or ""
+            translation="",
+            source=req.source,
+            target=req.target,
+            engine=req.engine or "",
         )
     if len(req.text) > MAX_INPUT_CHARS:
         raise HTTPException(400, f"Input too long (max {MAX_INPUT_CHARS} chars).")
@@ -87,19 +93,33 @@ def translate_endpoint(
 
     if provider.kind == "api":
         if not (provider.is_available() or api_key):
-            raise HTTPException(503, provider.setup_hint or "Add an API key in Settings.")
+            raise HTTPException(
+                503, provider.setup_hint or "Add an API key in Settings."
+            )
     elif not provider.is_available():
-        raise HTTPException(503, provider.setup_hint or f"Engine '{engine_id}' is not available.")
+        raise HTTPException(
+            503, provider.setup_hint or f"Engine '{engine_id}' is not available."
+        )
 
     try:
         result = provider.translate(req.text, req.source, req.target, api_key=api_key)
     except ValueError as exc:
         raise HTTPException(400, str(exc)) from exc
     except httpx.HTTPStatusError as exc:
-        raise HTTPException(502, f"Upstream API error: {exc.response.status_code}") from exc
+        raise HTTPException(
+            502, f"Upstream API error: {exc.response.status_code}"
+        ) from exc
     except Exception as exc:  # noqa: BLE001 — surface engine errors to the client
         raise HTTPException(500, f"Translation failed: {exc}") from exc
 
     return TranslateResponse(
         translation=result, source=req.source, target=req.target, engine=engine_id
     )
+
+
+@app.get("/scalar", include_in_schema=False)
+async def scalar_html():
+    return get_scalar_api_reference(
+        openapi_url=app.openapi_url,
+        title=app.title,
+    )

backend/providers/__init__.py

@@ -7,6 +7,8 @@ report themselves available.
 
 from __future__ import annotations
 
+import os
+
 from providers.base import ProviderInfo, TranslationProvider
 from providers.gemini import GeminiProvider
 from providers.groq import GroqProvider, GroqQwenProvider
@@ -14,14 +16,15 @@ from providers.madlad import MADLADProvider
 from providers.nllb import NLLBProvider
 from providers.ollama import OllamaProvider
 
-# Order = display order in the UI. Local/private engines first.
 _PROVIDERS: list[TranslationProvider] = [
     NLLBProvider(),
     NLLBProvider(
         provider_id="nllb_1_3b",
         name="NLLB-200 (1.3B)",
-        model_dir="models/nllb-200-distilled-1.3B-int8",
-        hf_model="facebook/nllb-200-distilled-1.3B",
+        model_dir=os.environ.get(
+            "CT2_MODEL_DIR_1_3B", "models/nllb-200-distilled-1.3B-int8"
+        ),
+        hf_model=os.environ.get("HF_MODEL_1_3B", "facebook/nllb-200-distilled-1.3B"),
     ),
     MADLADProvider(),
     OllamaProvider(),

backend/requirements.txt

@@ -12,3 +12,4 @@ protobuf==5.29.2
 
 # ctranslate2 4.5 imports pkg_resources, removed in setuptools>=81
 setuptools<81
+scalar-fastapi

entrypoint.sh

@@ -0,0 +1,19 @@
+#!/bin/bash
+set -e
+
+echo "== Converting NLLB 600M =="
+CT2_MODEL_DIR="/data/models/nllb-200-distilled-600M-int8" \
+python convert_model.py
+
+echo "== Converting NLLB 1.3B =="
+HF_MODEL="facebook/nllb-200-distilled-1.3B" \
+CT2_MODEL_DIR="/data/models/nllb-200-distilled-1.3B-int8" \
+python convert_model.py
+
+echo "== Converting MADLAD 3B =="
+HF_MODEL="google/madlad400-3b-mt" \
+CT2_MODEL_DIR="/data/models/madlad400-3b-mt-int8" \
+python convert_model.py
+
+echo "== Starting server =="
+exec uvicorn main:app --host 0.0.0.0 --port 7860