Skip to content

[None][infra] Add dev-container entrypoint dispatcher for CLI docker workflow#16574

Draft
jieli-matrix wants to merge 1 commit into
NVIDIA:mainfrom
jieli-matrix:feat/cli-docker-compose-dev-workflow
Draft

[None][infra] Add dev-container entrypoint dispatcher for CLI docker workflow#16574
jieli-matrix wants to merge 1 commit into
NVIDIA:mainfrom
jieli-matrix:feat/cli-docker-compose-dev-workflow

Conversation

@jieli-matrix

@jieli-matrix jieli-matrix commented Jul 19, 2026

Copy link
Copy Markdown
Collaborator

Why

The only supported container dev flow today is the VSCode Dev Container (.devcontainer/).
For a CLI-only developer it is a poor fit, and there
is no first-class path for the everyday bug-repro-and-fix loop: pull image → mount TOT +
model data → build (ccache, correct GPU arch) → pytest to repro → branch → incremental
rebuild → pytest to validate.

This PR adds a CLI-only, docker-compose develop-and-repro workflow that sits beside the
VSCode flow (which stays untouched), with a clean up -dexec loop and no generated env file.

Scope — where this runs, and what it does / doesn't cover

The two modes (rootful / rootless) select for docker daemon. Mapping to real environments:

Environment Container tech This workflow
Colossus / server station (interactive dev box) docker, real root (-u root:root, --privileged) rootful mode
Locked-down / shared HPC dev box (no root docker) rootless docker rootless mode
SLURM cluster (single- and multi-node srun jobs) enroot/pyxis (--container-remap-root) not a target — use examples/disaggregated/slurm/

Insights & analysis

  • The VSCode flow doesn't transfer to CLI. It depends on make_env.py generating
    .devcontainer/.env (run as VSCode's initializeCommand), and docker-compose.yml's command
    is just while sleep 1000 — a keep-alive with no build/test ergonomics.

  • up -d + exec beats run -it. Start a detached, named container once; drive builds/tests
    via exec.

TODO

  • .devcontainer/entrypoint.dev.sh — build/test dispatcher (build-full / build-cpp /
    build-python / build-precompiled / test / build-and-test / wait / shell). (first commit)
  • docker/compose.dev.yml (base: ipc: host, network_mode: host, init: true, shm_size
    1g/DEV_SHM_SIZE, ulimits.stack=67108864, caps — cluster runtime parity for single-node
    disagg) + docker/compose.dev.rootful.yml (adds user: + ulimit memlock=-1)
  • docker/dev wrapper — repo-local -f/--env-file; DEV_ROOTLESS=0|1 selector
  • Per-user config templates: docker/compose.dev.override.yml.example,
    docker/dev.user.env.example (git-ignore the real files)
  • Default CUDA arch → native in the entrypoint (fix the 90-real default)
  • Image-drift stamp (cpp/build/.built_with_image) → warn to build-full --clean on tag change
  • Developer-guide README for the CLI bug-repro loop

@coderabbitai summary

Description

Test Coverage

PR Checklist

Please review the following before submitting your PR:

  • PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.

  • PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.

  • Test cases are provided for new code paths (see test instructions)

  • If PR introduces API changes, an appropriate PR label is added - either api-compatible or api-breaking. For api-breaking, include BREAKING in the PR title.

  • Any new dependencies have been scanned for license and vulnerabilities

  • CODEOWNERS updated if ownership changes

  • Documentation updated as needed

  • Update tava architecture diagram if there is a significant design change in PR.

  • The reviewers assigned automatically/manually are appropriate for the PR.

  • Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.

…workflow

Introduce .devcontainer/entrypoint.dev.sh, a build/test dispatcher for a
CLI-only (non-VSCode) docker container dev loop: build-full, build-cpp,
build-python, build-precompiled, test, build-and-test, wait, shell.

All build state (ccache, cpp/build, symlinked .so) lives on the mounted
host tree, so the container stays disposable and rebuilds are incremental
across container/branch churn. This is the shared entrypoint for a
docker-compose based develop-and-repro workflow (WIP, draft).

Signed-off-by: Jie Li <lijie@nvidia.com>
@jieli-matrix jieli-matrix self-assigned this Jul 19, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant