vllm ocr

Dockerfile

@@ -0,0 +1,60 @@
+# Base: official vLLM OpenAI-compatible server (tested version family)
+FROM vllm/vllm-openai:v0.9.1
+
+# Work as root during build
+USER root
+
+# Speed up HF downloads and avoid interactive git prompts
+ENV HF_HUB_ENABLE_HF_TRANSFER=1 \
+    GIT_LFS_SKIP_SMUDGE=1 \
+    PIP_DISABLE_PIP_VERSION_CHECK=1
+
+# Install exact Transformers version dots.ocr expects (qwen2_vl APIs),
+# plus minimal utils. vLLM image already includes CUDA/PyTorch.
+RUN pip install --no-cache-dir "transformers==4.48.3" "huggingface_hub>=0.24.0"
+
+# Pre-download the model repo into /weights/DotsOCR
+# Note: dots.ocr requires the directory name to avoid '.' (see model card).
+ARG MODEL_ID=rednote-hilab/dots.ocr
+RUN python - <<'PY'
+from huggingface_hub import snapshot_download
+import os, shutil
+root = "/weights"
+os.makedirs(root, exist_ok=True)
+# Download into a temp dir then move everything under /weights/DotsOCR
+tmp = snapshot_download(repo_id=os.environ.get("MODEL_ID","rednote-hilab/dots.ocr"),
+                        local_dir=os.path.join(root,"_tmp"), allow_patterns=["*"])
+dst = os.path.join(root,"DotsOCR")
+os.makedirs(dst, exist_ok=True)
+for name in os.listdir(tmp):
+    if name == "DotsOCR": continue
+    shutil.move(os.path.join(tmp,name), os.path.join(dst,name))
+shutil.rmtree(tmp, ignore_errors=True)
+PY
+
+# Ensure the model code is importable as a top-level package "DotsOCR"
+ENV HF_MODEL_PATH=/weights/DotsOCR
+ENV PYTHONPATH="/weights:${PYTHONPATH}"
+
+# vLLM needs the custom model to be registered before main() runs.
+# The model authors recommend importing their vLLM adapter into the vllm CLI module.
+RUN sed -i '/^from vllm\.entrypoints\.cli\.main import main$/a\
+from DotsOCR import modeling_dots_ocr_vllm' $(which vllm)
+
+# Expose the Space port and launch the OpenAI-compatible server
+EXPOSE 7860
+ENV PORT=7860
+
+# Notes:
+# --chat-template-content-format string   per dots.ocr README
+# --served-model-name model               so clients can use model="model"
+# --trust-remote-code                     load repo’s custom code
+# --host 0.0.0.0 --port $PORT             bind to Space port
+CMD bash -lc '\
+  vllm serve ${HF_MODEL_PATH} \
+    --host 0.0.0.0 --port ${PORT} \
+    --served-model-name model \
+    --gpu-memory-utilization 0.95 \
+    --chat-template-content-format string \
+    --trust-remote-code \
+'