fix(docker): install smg-grpc-servicer from source in engine images

[key4ng]

Description

Problem

Engine images (ghcr.io/lightseekorg/smg:*-sglang-v0.5.10, etc.) shipped a stale smg_grpc_servicer inherited from whatever the base image (lmsysorg/sglang:v0.5.10) happened to preinstall from PyPI at its own build time — not the version in the SMG source tree being baked into the image.

When sglang v0.5.10 refactored srt/utils.py into the srt/utils/ subpackage and moved get_zmq_socket into srt/utils/network.py without re-exporting it at package level, the bundled stale servicer broke at import time:

File "/usr/local/lib/python3.12/dist-packages/smg_grpc_servicer/sglang/request_manager.py", line 38
    from sglang.srt.utils import get_or_create_event_loop, get_zmq_socket, kill_process_tree
ImportError: cannot import name 'get_zmq_socket' from 'sglang.srt.utils'
    (/sgl-workspace/sglang/python/sglang/srt/utils/__init__.py)

This crash-looped engine pods in production despite the v1.4.1 source tree already containing the correct import (commit 5321bce, released as smg-grpc-servicer 0.5.2 in #1078).

Why CI didn't catch it: scripts/ci_install_sglang.sh installs the servicer from source (uv pip install -e grpc_servicer/), so CI ran against the fixed code. The published Docker image ran against a stale PyPI snapshot baked into the base image. CI and the image were testing two different versions of the same package.

Solution

Eliminate the PyPI path entirely for engine images. Always install both smg-grpc-proto and smg-grpc-servicer from the cloned source tree, using --force-reinstall to override anything the base image preinstalled. This makes the image a single source of truth that tracks the repo.

Changes

• scripts/installation/install-smg.sh: after installing the smg wheel from bindings/python, also pip install --no-cache-dir --force-reinstall crates/grpc_client/python and grpc_servicer from the cloned source. No extras activated — engines are already in the base image and we don't want to reinstall vllm/sglang.
• docker/engine.Dockerfile: drop the now-redundant ENGINE=vllm-only pip install smg-grpc-servicer[vllm] block (covered uniformly by the source install).

Test Plan

Reproduction before this PR (from a crashing pod running ghcr.io/lightseekorg/smg:1.4.1-sglang-v0.5.10):

$ kubectl run smg-debug --image=ghcr.io/lightseekorg/smg:1.4.1-sglang-v0.5.10 \
    --restart=Never --command -- python3 -c \
    "from smg_grpc_servicer.sglang.servicer import SGLangSchedulerServicer"
...
ImportError: cannot import name 'get_zmq_socket' from 'sglang.srt.utils'

After this PR, install-smg.sh reinstalls smg-grpc-servicer from ${SMG_SRC}/grpc_servicer (which contains the fixed from sglang.srt.utils.network import get_zmq_socket) with --force-reinstall, overriding the base-image version.

Confirmation that the fix is actually live in the built image comes from the release-sglang-docker.yml PR dry-run on this branch: the initial build attempt (commit f091a01) successfully installed from source and got past request_manager.py:38 — the line the original get_zmq_socket ImportError surfaced at — and only failed deeper in the chain at sgl_kernel loading (which requires libcuda.so.1 and a real GPU driver, unavailable on a build runner). If the stale base-image servicer had still been active, the failure would have matched the production traceback exactly. That it didn't is proof --force-reinstall from source is in effect.

Note on defense-in-depth

I initially added a RUN python -c "from smg_grpc_servicer.sglang.servicer import ..." smoke test to the Dockerfile as defense-in-depth, but reverted it (commit 644980f) once CI proved it unrunnable: the servicer's import chain transitively loads sglang's sgl_kernel, which dlopens libcuda.so.1 at module init time and requires an NVIDIA driver that doesn't exist on CPU-only docker build runners. The appropriate place for a full import smoke test is a post-build job in release-sglang-docker.yml that runs the built image on a GPU runner — that's a separate workflow change and out of scope for this PR.

Local verification

• bash -n scripts/installation/install-smg.sh — syntax OK
• shellcheck scripts/installation/install-smg.sh — clean
• pre-commit run --files docker/engine.Dockerfile scripts/installation/install-smg.sh — all hooks pass
• cargo +nightly fmt --all -- --check — clean (Rust workspace unperturbed; no Rust files in diff)

Checklist

• cargo +nightly fmt passes (no-op — zero Rust changes)
• cargo clippy --all-targets --all-features -- -D warnings (no-op — zero Rust changes)
• (Optional) Documentation updated
• (Optional) Please join us on Slack #sig-smg to discuss, review, and merge PRs

Summary by CodeRabbit

• Build & Installation
  • Removed a previous special-case dependency install for the vLLM engine from the final Docker stage.
  • Added a forced reinstall of gRPC-related Python packages from source after building bindings to ensure installed packages match the current codebase and avoid stale versions.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro

Run ID: 6e065973-f696-4ca2-8b4e-8453d9383165

📥 Commits

Reviewing files that changed from the base of the PR and between f091a01 and 644980f.

📒 Files selected for processing (1)

• docker/engine.Dockerfile

💤 Files with no reviewable changes (1)

• docker/engine.Dockerfile

---

📝 Walkthrough

Walkthrough

The PR removes a vLLM-specific pip install step from docker/engine.Dockerfile and adds a forced pip reinstall of smg-grpc-proto and smg-grpc-servicer from source in scripts/installation/install-smg.sh. Engine install control flow and ENGINE validation remain unchanged.

Changes

Cohort / File(s)	Summary
Docker Engine Build docker/engine.Dockerfile	Removed the vLLM-specific pip install --no-cache-dir smg-grpc-proto "smg-grpc-servicer[vllm]" step in the final stage; left ENGINE validation and engine-source install script invocation unchanged.
SMG Installation Script scripts/installation/install-smg.sh	Added pip install --no-cache-dir --force-reinstall calls to install smg-grpc-proto and smg-grpc-servicer from local source paths after building Python wheels, ensuring repo source versions are installed.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~12 minutes

Possibly related PRs

• fix(docker): install gRPC proto and servicer for vLLM images #843: Related changes to installing smg-grpc-proto/smg-grpc-servicer for vLLM images; overlaps installation strategy.
• ci(docker): add engine-specific image build pipelines #651: Modifies the same Docker engine build and install script flow.
• refactor(docker): consolidate engine image build into reusable workflow #658: Introduced or modified the docker/engine build flow that this PR adjusts.

Suggested reviewers

• CatherineSue
• gongwei-130

Poem

🐰
Fresh wheels spun in morning light,
Reinstalling to make things right,
No stale bits left in the den,
Builds hop forward, try again—
A rabbit cheers: new source, new flight! 🥕✨

🚥 Pre-merge checks | ✅ 3

✅ Passed checks (3 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title directly describes the main change: installing smg-grpc-servicer from source in engine images, which is the core fix addressing the stale package issue documented in the PR objectives.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[gemini-code-assist]

Code Review

This pull request updates the Docker build and installation scripts to install smg-grpc-proto and smg-grpc-servicer directly from source, ensuring the container image remains synchronized with the repository. Additionally, a smoke-test step was added to the Dockerfile to verify successful imports of engine-specific servicers. Feedback suggests using python3 instead of python in the smoke-test to ensure compatibility across different base images.