Repository navigation
egocap: remove the capture server's GMSL frame-feed source - #1550
Conversation
The optional --gmsl-ipc mode made CaptureServer read camera frames directly from the receiver's CUDA IPC slots, JPEG-encode them and serve them as station cameras. Camera video reaches consumers through LiveKit, so the mode was a second direct reader of the GPU slots and extra surface to maintain. Removes the gmsl_ipc package, the source factory and the --gmsl-ipc, --gmsl-camera and --gmsl-fps options, their tests, and the gmsl dependency extra with its lock entries. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Signed-off-by: Gareth Morgan <gmorgan@nvidia.com>
|
Important Review skippedAuto incremental reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the
|
Automated reviewThe PR removes the optional GMSL frame-feed source (the gmsl_ipc package, the launch.py factory, the CLI options in server.py, the No issues found. ✅ Nothing blocking left from my side — auto-review of this one is now finished. 🤖 Sent by Shawn's Claude — automatic reply. Shawn hasn't reviewed this personally yet; treat it as a first pass. |
NVdejung
left a comment
There was a problem hiding this comment.
🤖 Automated first-pass review (Claude, for Dejun) — I'll follow up personally where needed.
Summary — This PR deletes the capture server's optional GMSL frame-feed source: the gmsl_ipc package and its tests, the GMSL factory in launch.py, the three --gmsl-* options, and the gmsl extra with its lock entries. The removal is complete. Nothing in this repository still imports the deleted modules or passes the removed flags or extra (the launcher never did, and sensing_gmsl is only an operator label, see operator_settings.py:45). Deployments outside the repository are the one thing I can't check. Two small suggestions, nothing blocking.
Must fix
None found. The internal-name scan of the added lines, title and description found nothing.
Suggestions
examples/egocap/uv.lock:254— The lock was edited by hand, anduv lock --checkwas never run. No checked-in CI job will catch a problem either: neither.github/workflows,.gitlab-ci.ymlnor the pre-commit config runs this tree's tests or checks its lock. So the "CI" box in the test plan tells you nothing aboutexamples/egocap. I read the lock at head. It removes exactly what uv would: thecupy-cuda12xandfastrlockblocks, thegmsloptional-dependency group, itsrequires-distrow and theprovides-extrasentry.fastrlockhad no other dependent, and removing a package only drops constraints, so a re-lock should leave the file unchanged. Still, please runuv lock --check(or regenerate the lock) with a uv that parses the project before merging. It takes one command and settles the unchecked box.examples/egocap/capture_server/capture_server/server.py:2194— The comment no longer mentionsFeedError, but theexceptstill catchesRuntimeError, which the old comment said was there for the feed. Nothing in thetryblock raises it directly any more:load_episoderaisesEpisodeFormatError/FileNotFoundError, and the remainingraise RuntimeErrorsites all run during a session. Either drop it, or keep it on purpose (for example, for zarr or OpenCV errors during episode load) and say so in the comment. Otherwise an unexpected setupRuntimeErrorprints as a one-linecapture-server:error with exit 2 instead of a traceback. Low priority.
Conflicts
- With
main: 0 commits behind (the merge-base ismain's head), and the local 3-way merge is clean, so no rebase is needed. GitHub'sblockedstate is branch protection, not a conflict. - With open PRs:
#1551,#1552and#1553(dependabot bumps of opentelemetry-api, certifi and pillow) also changeexamples/egocap/uv.lock. There is no textual conflict, because their hunks are nowhere near the removed blocks. There is no semantic one either: none of those packages depends on cupy or fastrlock. They can merge in any order. If dependabot rebases one of them after this PR lands, it regenerates the lock from the newpyproject.toml, which also independently checks this hand edit.
Change-by-change (10 files)
Paths are relative to examples/egocap/. I read the deleted files in full to confirm nothing else depends on them.
| File / area | What changed | Makes sense? |
|---|---|---|
capture_server/README.md (L1253) |
LiveKit name-clash rule no longer tied to --gmsl-camera |
✅ matches the check in server.py:209 (cameras plus known_cameras), tested at test_livekit_streams.py:246 |
capture_server/capture_server/gmsl_ipc/ (__init__.py, feed.py, cuda_slots.py, source.py) |
Package deleted: feed client, CUDA IPC slot mapper, frame-feed Pipeline |
✅ nothing in the repo imports it any more |
capture_server/capture_server/launch.py (L27) |
Removes GmslCamera, gmsl_station_cameras, build_gmsl_source_factory, _FeedAcquisition, the dataclass and type-checking imports, and their __all__ entries |
✅ every remaining import is still used |
capture_server/capture_server/server.py CLI (L1993, L2217) |
Removes --gmsl-ipc (from the source group), --gmsl-camera, --gmsl-fps, _gmsl_cameras, _gmsl_setup, the _run_cli branch and the two main checks; rewords the --camera error |
✅ --camera with --mock or --replay is still rejected and tested (test_cli.py:13, test_replay.py:390); the untouched group comment ("with neither flag…") is accurate again |
capture_server/capture_server/server.py _run_cli (L2194) |
Setup-error comment no longer mentions FeedError |
try block raises RuntimeError directly any more (see suggestions) |
pyproject.toml (L59) |
gmsl extra removed |
✅ |
uv.lock (L254) |
cupy-cuda12x, fastrlock and the gmsl extra metadata removed by hand |
|
tests/capture_server/test_gmsl_ipc.py |
Deleted (21 tests) | ✅ tested only the deleted code; its shared helpers still have other users |
Reviewed at 685bbc5 · 10 files · 2026-11-18 14:39 PST
NVshawnhwang
left a comment
There was a problem hiding this comment.
Reviewed at 685bbc5. No actionable findings.
The GMSL frame-feed removal is complete: no remaining references to the removed API or flags were found under examples/egocap, and the launcher continues to use the unchanged LiveKit path. The deleted tests cover the removed functionality.
Independently validated locally:
- Capture-server and tools suites: 899 passed, 2 skipped.
uv lock --check --offlinewith uv 0.13.7: passed, closing the lockfile validation gap noted in the description.- Ruff lint and formatting checks on the changed Python files: passed.
- CLI help and parsing:
--mockparses and--gmsl-ipcis rejected as intended.
External deployments that still use --gmsl-ipc remain outside the scope of repository verification.
Reviewed with Codex.
The setup except clause caught RuntimeError for the removed GMSL frame feed. Nothing in setup raises it directly any more, so an unexpected one now shows a traceback instead of a one-line error. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Signed-off-by: Gareth Morgan <gmorgan@nvidia.com>
|
@NVdejung thanks for the review. Both suggestions are handled:
🤖 Generated with Claude Code |
Summary
Removes the capture server's optional GMSL frame-feed camera source. The
--gmsl-ipcmode made CaptureServer read camera frames directly from the receiver's CUDA IPC slots, JPEG-encode them and serve them as station cameras. Camera video reaches consumers through LiveKit and the launcher never used this mode, so it was a second direct reader of the receiver's GPU slots and extra surface to maintain.What changed
gmsl_ipcpackage (feed client, CUDA slot mapper, camera source) and its tests.launch.pyand the--gmsl-ipc,--gmsl-cameraand--gmsl-fpsoptions and their setup fromserver.py.gmsloptional-dependency extra frompyproject.tomland the cupy and fastrlock entries fromuv.lock; nothing else depended on them.--gmsl-camerareference and keeps the rule that a LiveKit stream name may not equal a configured camera name.sensing_gmslcamera type in operator settings is a different thing and is unchanged.Validation
tests/capture_server: 779 passed, 0 failed (the 21 fewer than before are the removed GMSL tests); capture_server coverage 93%.tests/tools: 120 passed, 2 skipped (those need tools from other PRs).uv.lockandpyproject.tomlparse; no cupy, fastrlock or gmsl references remain, and no repo references to the removed names remain.capture-server --helphas no GMSL options and--mockstill parses;--gmsl-ipcis now rejected as an unrecognised argument.uv lock --checkwas not run: the installed uv cannot parse the project file, so the lock file was edited by hand and checked by parsing and grep only.Known limitations / follow-ups
capture-server --gmsl-ipcloses that mode; the launcher and the LiveKit bridge path are unaffected.Test plan
--gmsl-ipc, and regenerate the lock with a currentuvif preferred.🤖 Generated with Claude Code
Summary by CodeRabbit
--gmsl-camera.