Update Dockerfile

Dockerfile

@@ -7,40 +7,46 @@ ENV PIP_NO_CACHE_DIR=1 \
     PYTHONUNBUFFERED=1 \
     EMBED_MODEL=Seznam/simcse-small-e-czech \
     MIN_TOPIC_SIZE=10 \
-    MAX_DOCS=5000
-
-# ------------------------------------------------------------------
-# 2. Writable cache directories for Hugging Face & sentence-transformers
-# ------------------------------------------------------------------
-ENV HF_HOME=/tmp/hfcache \
+    MAX_DOCS=5000 \
+    HF_HOME=/tmp/hfcache \
     TRANSFORMERS_CACHE=/tmp/hfcache \
-    SENTENCE_TRANSFORMERS_HOME=/tmp/hfcache
-RUN mkdir -p /tmp/hfcache \
- && chmod -R 777 /tmp/hfcache
+    SENTENCE_TRANSFORMERS_HOME=/tmp/hfcache \
+    NUMBA_CACHE_DIR=/tmp/numba_cache
 
 # ------------------------------------------------------------------
-# 3. Writable cache dir for numba (already used in app.py)
+# 2. Make caches and give them world‐writable perms
 # ------------------------------------------------------------------
-ENV NUMBA_CACHE_DIR=/tmp/numba_cache
-RUN mkdir -p /tmp/numba_cache \
- && chmod -R 777 /tmp/numba_cache
+RUN mkdir -p /tmp/hfcache /tmp/numba_cache \
+ && chmod -R 777 /tmp/hfcache /tmp/numba_cache
 
 # ------------------------------------------------------------------
-# 4. Install Python deps and copy code
+# 3. Install Python deps
 # ------------------------------------------------------------------
 WORKDIR /code
 COPY requirements.txt .
 RUN pip install --no-cache-dir -r requirements.txt
 
-# ---- PRE-DOWNLOAD Czech SBERT so runtime never pulls ----
+# ------------------------------------------------------------------
+# 4. Pre-download the model/tokenizer into HF cache
+#    Using transformers directly to populate both
+#    – the “models--…” layout that HF v4+ uses
+#    – and then we’ll symlink so SBERT’s fallback path works too
+# ------------------------------------------------------------------
 RUN python - <<'PY'
-from sentence_transformers import SentenceTransformer
-SentenceTransformer(
-    'Seznam/simcse-small-e-czech',
-    cache_folder='/tmp/hfcache'
-)
+from transformers import AutoTokenizer, AutoModel
+AutoTokenizer.from_pretrained("Seznam/simcse-small-e-czech")
+AutoModel.from_pretrained("Seznam/simcse-small-e-czech")
 PY
 
+# ------------------------------------------------------------------
+# 5. Symlink the two possible cache‐dirs so SBERT never misses it
+# ------------------------------------------------------------------
+RUN ln -s /tmp/hfcache/models--Seznam--simcse-small-e-czech /tmp/hfcache/Seznam_simcse-small-e-czech \
+ && chmod -R 777 /tmp/hfcache/models--Seznam--simcse-small-e-czech /tmp/hfcache/Seznam_simcse-small-e-czech
+
+# ------------------------------------------------------------------
+# 6. Copy your app and expose
+# ------------------------------------------------------------------
 COPY app.py .
 
 EXPOSE 7860