.woodpecker.yml

@@ -5,14 +5,10 @@
 # translation happens here, asynchronously, and every generated sibling lands
 # as a reviewable bot commit.
 #
-# Translation is self-hosted — OPUS-MT runs on CPU inside the image, so there
+# Translation is self-hosted — the model ships inside vienalatina/translate,
-# is no API key and no third-party request. Before the first run, on the server:
+# so there is no API key and no third-party request. Build that image on the
-#   bash scripts/fetch-models.sh                        # -> /srv/mt-models
+# server before the first run:
-#   docker build -t vienalatina/translate:2 docker/translate
+#   docker build -t vienalatina/translate:1 docker/translate
-#
-# The models are mounted rather than baked in, so swapping them later does not
-# mean rebuilding the image. The repo must be marked "trusted" in Woodpecker
-# for this mount, which it already is for the deploy step.
 #
 # Secrets to configure in Woodpecker (repo settings → secrets):
 #   gitea_push_token — Gitea token for the translations bot user
@@ -27,10 +23,8 @@ when:
 
 steps:
   translate:
-    image: vienalatina/translate:2
+    image: vienalatina/translate:1
     pull: false          # built locally on the server, never fetched from a registry
-    volumes:
-      - /srv/mt-models:/opt/mt/models:ro
     environment:
       GITEA_PUSH_TOKEN:
         from_secret: gitea_push_token

content/page/acerca.de.md

@@ -0,0 +1,8 @@
+---
+title: "Über uns"
+lang: de
+manual_translation: true
+translated_from: es
+---
+
+**Viena Latina** ist die lateinamerikanische Gemeinschaft in Wien: Tourismusführer, kulturelle Agenda, Gastronomie, Gemeinschaftsleben und lateinamerikanische Geschäfte in der Stadt.

content/page/acerca.pt-br.md

@@ -0,0 +1,8 @@
+---
+title: Sobre
+lang: pt-br
+manual_translation: false
+translated_from: es
+---
+
+**Viena Latina** é a comunidade latino-americana em Viena: guias de turismo, agenda cultural, gastronomia, vida comunitária e lojas latinas na cidade.

content/page/contacto.de.md

@@ -0,0 +1,8 @@
+---
+title: Kontakt
+lang: de
+manual_translation: false
+translated_from: es
+---
+
+Schreiben Sie uns an **hola@vienalatina.com**.

content/page/contacto.pt-br.md

@@ -0,0 +1,8 @@
+---
+title: Contacto
+lang: pt-br
+manual_translation: false
+translated_from: es
+---
+
+Escreva-nos a **hola@vienalatina.com**.

content/post/bienvenida-a-viena-latina.de.md

@@ -0,0 +1,18 @@
+---
+title: "Willkommen bei Viena Latina"
+date: 2026-07-31
+slug: bienvenida-a-viena-latina
+lang: de
+manual_translation: false
+categories: [Comunidad]
+description: "Die neue Website von Viena Latina: schneller, selbst gehostet und mit überprüfbaren Übersetzungen auf Deutsch und Portugiesisch."
+---
+
+Das ist das neue Zuhause von **Viena Latina**. Die Website ist jetzt komplett
+statisch: Jeder Artikel wird auf Spanisch veröffentlicht und automatisch ins
+Deutsche und brasilianische Portugiesisch übersetzt, mit eigenen URLs pro
+Sprache.
+
+Für unsere Leserinnen und Leser ändert sich nichts — dieselben Rubriken wie
+immer: Tourismus, Kultur, Gastronomie, Gemeinschaft und Handel. Alles lädt
+schneller, ohne Cookies und ohne Tracker.

content/post/bienvenida-a-viena-latina.es.md

@@ -0,0 +1,16 @@
+---
+title: "Bienvenida a Viena Latina"
+date: 2026-07-31
+lang: es
+manual_translation: false
+categories: [Comunidad]
+description: "El nuevo sitio de Viena Latina: más rápido, autoalojado y con traducciones revisables en alemán y portugués."
+---
+
+Este es el nuevo hogar de **Viena Latina**. El sitio ahora es completamente
+estático: cada artículo se publica en español y se traduce automáticamente al
+alemán y al portugués brasileño, con URLs propias para cada idioma.
+
+Nada cambia para quienes nos leen — las mismas secciones de siempre: Turismo,
+Cultura, Gastronomía, Comunidad y Comercio. Todo carga más rápido, sin cookies
+y sin rastreadores.

content/post/bienvenida-a-viena-latina.pt-br.md

@@ -0,0 +1,17 @@
+---
+title: "Bem-vindos ao Viena Latina"
+date: 2026-07-31
+slug: bienvenida-a-viena-latina
+lang: pt-br
+manual_translation: false
+categories: [Comunidad]
+description: "O novo site do Viena Latina: mais rápido, auto-hospedado e com traduções revisáveis em alemão e português."
+---
+
+Este é o novo lar do **Viena Latina**. O site agora é totalmente estático:
+cada artigo é publicado em espanhol e traduzido automaticamente para o alemão
+e o português brasileiro, com URLs próprias para cada idioma.
+
+Nada muda para quem nos lê — as mesmas seções de sempre: Turismo, Cultura,
+Gastronomia, Comunidade e Comércio. Tudo carrega mais rápido, sem cookies e
+sem rastreadores.

content/post/hola-mundo.es.md

@@ -1,13 +0,0 @@
----
-title: "Hola mundo"
-date: 2026-09-18
-lang: es
-manual_translation: false
-categories: [Comunidad]
----
-
-Bienvenidos al nuevo sitio de Viena Latina.
-
-Publicamos en español, y cada artículo aparece también en alemán y en portugués.
-Aquí vas a encontrar turismo, cultura, gastronomía, comunidad y comercio latino
-en Viena.

docker/translate/Dockerfile

@@ -1,20 +1,21 @@
 # Pipeline image for the translate step.
 #
-# Models are NOT baked in — they live on the host at /srv/mt-models (see
+# Model and tokenizer are baked in, so a publish makes zero network calls and
-# scripts/fetch-models.sh) and are mounted read-only by .woodpecker.yml. That
+# needs no API key. torch is deliberately absent: it is only required to
-# keeps this image small and lets a model change be a directory swap instead of
+# *convert* a model, and the default PyPI wheel drags in ~2.5GB of CUDA
-# a 1GB image rebuild, which matters because model choice turned out to need
+# libraries this CPU-only box will never use.
-# iteration.
 #
-# torch is deliberately absent: it is only needed to *convert* a model, and the
+# Build on the server (once, and again only when changing models):
-# default PyPI wheel drags in ~2.5GB of CUDA libraries this CPU-only box will
+#   docker build -t vienalatina/translate:1 docker/translate
-# never use. Conversion happens in fetch-models.sh, not here.
 #
-# Build on the server:
+# The default model is a pre-converted CTranslate2 build, which avoids running
-#   docker build -t vienalatina/translate:2 docker/translate
+# ct2-transformers-converter on a 4GB box — it gets OOM-killed there.
 
 FROM python:3.12-slim
 
+ARG MT_MODEL_REPO=michaelfeil/ct2fast-m2m100_418M
+ARG MT_TOKENIZER_REPO=facebook/m2m100_418M
+
 # translate.py shells out to git to diff the push and commit the siblings back.
 RUN apt-get update -qq \
  && apt-get install -qq -y --no-install-recommends git \
@@ -25,9 +26,17 @@ RUN pip install --no-cache-dir \
       transformers \
       sentencepiece \
       sentencex \
-      pyyaml
+      pyyaml \
+      huggingface_hub
 
-ENV MT_PROVIDER=opus \
+RUN python -c "from huggingface_hub import snapshot_download; \
-    MT_MODEL_DIR=/opt/mt/models \
+snapshot_download('${MT_MODEL_REPO}', local_dir='/opt/mt/model')" \
+ && python -c "import transformers; \
+transformers.AutoTokenizer.from_pretrained('${MT_TOKENIZER_REPO}') \
+  .save_pretrained('/opt/mt/tokenizer')"
+
+# The published artifact is float16; CTranslate2 quantises to int8 on load.
+ENV MT_MODEL_DIR=/opt/mt/model \
+    MT_TOKENIZER=/opt/mt/tokenizer \
     MT_COMPUTE_TYPE=int8 \
     MT_THREADS=2

fetch-models.sh

@@ -1,30 +0,0 @@
-#!/usr/bin/env bash
-# Download and convert the OPUS-MT models into $DEST (default /srv/mt-models).
-# Containers run as root so pip lands on PATH and can write to the root-owned
-# target; the pipeline mounts that directory read-only.
-set -euo pipefail
-DEST="${1:-/srv/mt-models}"
-MODELS=(
-  "Helsinki-NLP/opus-mt-es-de"
-  "Helsinki-NLP/opus-mt-tc-big-de-es"
-  "Helsinki-NLP/opus-mt-tc-big-itc-itc"
-)
-for REPO in "${MODELS[@]}"; do
-  NAME="${REPO##*/}"
-  if [ -f "$DEST/$NAME/model.bin" ]; then
-    echo "== $NAME already present, skipping"
-    continue
-  fi
-  echo "== converting $NAME"
-  docker run --rm --memory=3g -v "$DEST:/out" python:3.12-slim bash -c "
-    set -e
-    pip install --quiet 'transformers[torch]' ctranslate2 sentencepiece \
-      --extra-index-url https://download.pytorch.org/whl/cpu
-    ct2-transformers-converter --model $REPO --output_dir /out/$NAME --quantization int8
-    python -c 'import sys
-from transformers import AutoTokenizer
-AutoTokenizer.from_pretrained(sys.argv[1]).save_pretrained(sys.argv[2])' $REPO /out/$NAME
-  "
-done
-echo
-du -sh "$DEST"/*

scripts/fetch-models.sh

@@ -1,46 +0,0 @@
-#!/usr/bin/env bash
-# Download and convert the OPUS-MT models the pipeline needs.
-#
-# Run once on the server; the pipeline mounts the result read-only, so changing
-# models later means re-running this rather than rebuilding a 1GB image.
-#
-#   bash scripts/fetch-models.sh            # -> /srv/mt-models
-#   bash scripts/fetch-models.sh /tmp/mt    # -> somewhere else
-#
-# Conversion holds the fp32 model in RAM. These are 74M-237M parameter models,
-# so each peaks well under the 4GB box's headroom — unlike M2M100 418M, whose
-# conversion was OOM-killed here at a 3GB cap. One container per model keeps a
-# failure contained to that model.
-
-set -euo pipefail
-DEST="${1:-/srv/mt-models}"
-
-MODELS=(
-  "Helsinki-NLP/opus-mt-es-de"
-  "Helsinki-NLP/opus-mt-tc-big-de-es"
-  "Helsinki-NLP/opus-mt-tc-big-itc-itc"
-)
-
-sudo mkdir -p "$DEST"
-sudo chown "$(id -u):$(id -g)" "$DEST"
-
-for REPO in "${MODELS[@]}"; do
-  NAME="${REPO##*/}"
-  if [ -f "$DEST/$NAME/model.bin" ]; then
-    echo "== $NAME already converted, skipping"
-    continue
-  fi
-  echo "== converting $NAME"
-  docker run --rm -u "$(id -u):$(id -g)" -e HOME=/tmp --memory=3g \
-    -v "$DEST:/out" python:3.12-slim bash -c "
-      set -e
-      pip install --quiet 'transformers[torch]' ctranslate2 sentencepiece \
-        --extra-index-url https://download.pytorch.org/whl/cpu
-      ct2-transformers-converter --model $REPO --output_dir /out/$NAME \
-        --quantization int8 --copy_files source.spm target.spm vocab.json tokenizer_config.json
-    "
-done
-
-echo
-du -sh "$DEST"/*
-echo "Models ready in $DEST"

scripts/translate.py

@@ -37,18 +37,10 @@ from pathlib import Path
 
 import yaml
 
+from translation.ctranslate_provider import CTranslate2Provider
 from translation.markdown import translate_markdown, translate_text
 from translation.provider import SITE_LANGS
 
-
-def build_provider():
-    """MT_PROVIDER=m2m100 falls back to the single-model engine."""
-    if os.environ.get("MT_PROVIDER", "opus") == "m2m100":
-        from translation.ctranslate_provider import CTranslate2Provider
-        return CTranslate2Provider()
-    from translation.opus_provider import OpusMTProvider
-    return OpusMTProvider()
-
 REPO_ROOT = Path(__file__).resolve().parent.parent
 CONTENT_DIR = REPO_ROOT / "content"
 
@@ -233,7 +225,7 @@ def main() -> None:
         print("No authored content changed — nothing to translate.")
         return
 
-    provider = build_provider()
+    provider = CTranslate2Provider()
     written: list[Path] = []
     for source, basename, lang in sources:
         print(f"Translating {source.relative_to(REPO_ROOT)} (from {lang}):")

scripts/translation/opus_provider.py

@@ -1,83 +0,0 @@
-"""OPUS-MT (Helsinki-NLP) via CTranslate2, on CPU.
-
-Replaces M2M100 418M, which was measured producing unusable German on real
-articles — "frijoles" became "Beeren" (berries), "rompe la idea" became
-"breitet die Idee" (spreads it), and the site's own name came back mangled.
-
-Not every pair exists as a published model, so routes are explicit rather than
-assumed. Verified against Hugging Face:
-
-    es -> de        opus-mt-es-de            (small, ~74M)
-    de -> es        opus-mt-tc-big-de-es     (tc-big, ~237M)
-    es <-> pt-br    opus-mt-tc-big-itc-itc   (Italic multilingual)
-    de <-> pt-br    no direct model — pivots through Spanish
-
-Marian models take the target language as a `>>xxx<<` token at the start of
-the *source* text, unlike M2M100's decoder-side target_prefix.
-"""
-
-from __future__ import annotations
-
-import os
-
-from .provider import Provider
-
-# (source, target) -> [(model directory, target-language token or None), ...]
-# More than one hop means a pivot translation.
-ROUTES: dict[tuple[str, str], list[tuple[str, str | None]]] = {
-    ("es", "de"): [("opus-mt-es-de", None)],
-    ("de", "es"): [("opus-mt-tc-big-de-es", None)],
-    ("es", "pt-br"): [("opus-mt-tc-big-itc-itc", ">>pob<<")],
-    ("pt-br", "es"): [("opus-mt-tc-big-itc-itc", ">>spa<<")],
-    ("de", "pt-br"): [("opus-mt-tc-big-de-es", None),
-                      ("opus-mt-tc-big-itc-itc", ">>pob<<")],
-    ("pt-br", "de"): [("opus-mt-tc-big-itc-itc", ">>spa<<"),
-                      ("opus-mt-es-de", None)],
-}
-
-
-class OpusMTProvider(Provider):
-    def __init__(self, model_root: str | None = None, compute_type: str | None = None,
-                 threads: int | None = None):
-        self.model_root = model_root or os.environ.get("MT_MODEL_DIR", "/opt/mt/models")
-        self.compute_type = compute_type or os.environ.get("MT_COMPUTE_TYPE", "int8")
-        self.threads = threads or int(os.environ.get("MT_THREADS", "2"))
-        self._name: str | None = None
-        self._translator = None
-        self._tokenizer = None
-
-    def _load(self, name: str):
-        """Keep exactly one model resident — three at once would not fit a 4GB box."""
-        if self._name == name:
-            return self._translator, self._tokenizer
-        import ctranslate2
-        import transformers
-
-        self._translator = None  # free the previous model before allocating the next
-        path = os.path.join(self.model_root, name)
-        self._tokenizer = transformers.AutoTokenizer.from_pretrained(path)
-        self._translator = ctranslate2.Translator(
-            path, device="cpu", compute_type=self.compute_type, intra_threads=self.threads
-        )
-        self._name = name
-        return self._translator, self._tokenizer
-
-    def _hop(self, texts: list[str], model: str, token: str | None) -> list[str]:
-        translator, tok = self._load(model)
-        prepared = [f"{token} {t}" if token else t for t in texts]
-        batch = [tok.convert_ids_to_tokens(tok.encode(t)) for t in prepared]
-        results = translator.translate_batch(batch)
-        return [
-            tok.decode(tok.convert_tokens_to_ids(r.hypotheses[0]), skip_special_tokens=True)
-            for r in results
-        ]
-
-    def translate(self, texts: list[str], src: str, tgt: str) -> list[str]:
-        if not texts:
-            return []
-        route = ROUTES.get((src, tgt))
-        if route is None:
-            raise ValueError(f"no translation route from {src} to {tgt}")
-        for model, token in route:
-            texts = self._hop(texts, model, token)
-        return texts