Improve deployment setup: requirements.txt, install.sh, Dockerfile, and README updates

[Hardik-369]

Closes #27

This PR addresses several of the deployment blockers raised in #27.

Changes

• requirements.txt — Pinned dependencies for both Transformers and SGLang inference paths. Uses --find-links wheel to resolve the custom SGLang wheel.
• install.sh — Automated uv-based environment creation and dependency installation. Includes a prominent warning that the custom wheel is required.
• Dockerfile — Container image based on nvidia/cuda:12.9.0-runtime-ubuntu24.04 with Python 3.12, the custom SGLang wheel, and all dependencies.
• README updates — Added a prominent warning about the custom SGLang wheel, a Quick Start section, and Docker usage examples.

[Copilot]

Pull request overview

This PR improves deployment reproducibility and onboarding by adding pinned dependency installation, an automated setup script, containerization via Docker, and clearer README setup/run instructions—addressing the deployment blockers described in #27 (especially the custom SGLang wheel and lack of Docker support).

Changes:

• Added a pinned requirements.txt that points pip to the bundled wheel/ directory for the custom SGLang wheel.
• Introduced install.sh to automate environment creation and dependency installation using uv.
• Added a CUDA 12.9 + Python 3.12 Dockerfile and expanded README with prerequisites, quick start, and Docker usage.

Reviewed changes

Copilot reviewed 3 out of 4 changed files in this pull request and generated 5 comments.

File	Description
requirements.txt	Adds a dependency set intended to be reproducible and to resolve the bundled SGLang wheel via --find-links.
install.sh	Automates venv creation + installs the bundled SGLang wheel and requirements using uv.
Dockerfile	Provides a CUDA runtime image that installs the custom wheel + dependencies and launches an SGLang server by default.
README.md	Documents prerequisites, quick start installation paths, and Docker build/run examples.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[kushdab]

Solid direction -- the install.sh + requirements.txt combo is exactly what new users need, and the README warning about the custom SGLang wheel will save a lot of confusion. A few issues worth fixing before this lands:

Bugs

1. --trust-remote-code missing from Docker CMD

baidu/Unlimited-OCR ships custom model code in modeling_unlimitedocr.py. Without this flag, SGLang refuses to load it and exits immediately:

Error: Loading baidu/Unlimited-OCR requires --trust-remote-code

Add it to the CMD:

CMD ["python", "-m", "sglang.launch_server", \
    "--model", "baidu/Unlimited-OCR", \
    "--trust-remote-code", \          # <- add this
    ...

2. kernels==0.11.7 in requirements.txt causes the SGLang ValueError

The custom SGLang wheel bundles its own sgl_kernel. Installing kernels==0.11.7 alongside it downgrades sgl_kernel, which breaks SGLang's unlimited_ocr.py model registration at import time -- SGLang swallows the import error, starts anyway, then crashes at inference with ValueError: UnlimitedOCRForCausalLM is not supported by SGLang (the root cause of issue #12). PR #34 removes this line from the README; the same fix applies here:

-kernels==0.11.7

This affects all three install paths: install.sh, direct pip install -r requirements.txt, and the Dockerfile.

3. --image-dir / --output-dir in README Docker examples should be --image_dir / --output_dir

infer.py uses underscore-delimited argparse flags. The Docker README examples use hyphens, which will produce unrecognized arguments errors:

# As written (broken):
docker run ... unlimited-ocr python infer.py --image-dir /data --output-dir /app/outputs

# Fix:
docker run ... unlimited-ocr python infer.py --image_dir /data --output_dir /app/outputs

Security / image quality

4. Container runs as root

The Dockerfile has no useradd/USER directive, so the container runs as root. This is a security concern for production deployments. Adding a non-root user is straightforward:

RUN groupadd --gid 1000 ocr && \
    useradd --uid 1000 --gid 1000 --create-home --shell /bin/bash ocr && \
    chown -R ocr:ocr /app
USER ocr

5. COPY assets/ assets/ adds documentation images to the runtime image

The assets/ folder contains the README GIF and other images used only for documentation. Copying them into the container adds unnecessary size. Either drop the COPY assets/ line (nothing in infer.py reads from assets/ at runtime) or add a .dockerignore to keep the build context clean.

A minimal .dockerignore:

.git/
__pycache__/
*.py[cod]
.venv/
outputs/
log/
assets/
*.pdf

Comparison with PR #39

Both PRs address the same deployment need (#27). Key differences:

	PR #39 (marcelMaier)	PR #41 (this PR)
Base image	devel (wrong, ~7.5 GB)	runtime (correct, ~3.5 GB)
Multi-stage build	Yes	No
Non-root user	Yes	No
install.sh	No	Yes
requirements.txt	No	Yes
--trust-remote-code in CMD	Missing	Missing
kernels==0.11.7	Present (wrong)	Present (wrong)

This PR's nvidia/cuda:12.9.0-runtime-ubuntu24.04 base is the correct choice (PR #39 mistakenly uses devel). If the maintainers want to merge one Docker implementation, cherry-picking the runtime base, install.sh, and requirements.txt from this PR while adopting the multi-stage build and non-root user from PR #39 would give the best combined result.

What's done well

• nvidia/cuda:12.9.0-runtime-ubuntu24.04 (runtime, not devel) is the right base -- cuts final image size roughly in half vs. PR Add Docker image publishing and API runtime #39
• Layer ordering is cache-optimized: wheel first (changes rarely), requirements second, application code last
• --find-links wheel in requirements.txt cleanly handles the local wheel without manual path instructions
• install.sh is the clearest new-user onboarding addition in any open PR right now -- the prominent warning about the custom wheel alone will close several support issues
• Quick Start section in README is well-structured and covers both uv and manual pip paths