Upload Dockerfile with huggingface_hub

Dockerfile

@@ -1,26 +1,12 @@
-FROM ghcr.io/huggingface/text-generation-inference:latest
+FROM vllm/vllm-openai:latest
 
-# Secrets injected at runtime via HF Space settings:
-#   HF_TOKEN     — HF token with google/gemma-4-E2B-it access
-#   VLLM_API_KEY — Bearer token clients must send
+# Gemma 4 (model_type=gemma4) needs newer transformers than vLLM bundles.
+# TGI does NOT support gemma4 yet — this image uses vLLM which does.
+RUN pip install --no-cache-dir "transformers>=4.51.0" "accelerate>=0.34.0"
+
+COPY entrypoint.sh /entrypoint.sh
+RUN chmod +x /entrypoint.sh
 
 EXPOSE 7860
 
-# TGI reads HUGGING_FACE_HUB_TOKEN for model download auth.
-# API_KEY is TGI's native env var for request authentication.
-# --trust-remote-code ensures Gemma 4 chat template + architecture loads correctly.
-# --max-input-length / --max-total-tokens set 128k context window.
-# --max-concurrent-requests 8 — serves up to 8 simultaneous episodes.
-ENTRYPOINT []
-CMD ["sh", "-c", "\
-  HUGGING_FACE_HUB_TOKEN=$HF_TOKEN \
-  API_KEY=$VLLM_API_KEY \
-  text-generation-launcher \
-    --model-id google/gemma-4-E2B-it \
-    --port 7860 \
-    --dtype bfloat16 \
-    --max-input-length 126000 \
-    --max-total-tokens 131072 \
-    --max-concurrent-requests 8 \
-    --trust-remote-code \
-"]
+ENTRYPOINT ["/entrypoint.sh"]