CI: add DSpark 16-card serving accuracy guard

[ndleslx]

• Add a separate serving-dspark PR job for changes under
  models/deepseek_v4_flash_dspark, with its own build and pip caches.
• Run the Palace 64-prompt/128-completion HTTP guard from
  pypto-serving/main on 16 cards against the pypto-lib checkout under
  test. Print completions for review and check the response contract
  and token accounting with default Simpler timeouts.
• Resolve PYPTO_DSV4_DSPARK_MODEL_DIR from runner configuration,
  validate the weights directory, and forward it through task-submit.
• Bound queue and execution time and retain orphan-task and build
  artifact cleanup for the new serving job.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Advanced

Run ID: 8f476527-eea5-4b8e-b68b-8c4366619ca8

📥 Commits

Reviewing files that changed from the base of the PR and between aff70c1 and 2494855.

📒 Files selected for processing (3)

• .github/actions/pypto-serving-tests/action.yml
• .github/actions/setup-ci-job/action.yml
• .github/workflows/ci.yml

Included review availability: Your plan provides up to 1 included review per hour; 0 remain after this review.

---

📝 Walkthrough

Walkthrough

The CI detects changes to the DeepSeek V4 DSpark model and runs a dedicated serving accuracy job. The serving action resolves the DSpark model directory and executes tests on 16 devices.

Changes

DeepSeek DSpark serving tests

Layer / File(s)	Summary
DSpark serving action .github/actions/pypto-serving-tests/action.yml, .github/actions/setup-ci-job/action.yml	The serving action accepts deepseek-v4-dspark, resolves PYPTO_DSV4_DSPARK_MODEL_DIR, and runs the DSpark accuracy test on 16 devices. Runner configuration documentation includes the new environment variable.
Workflow detection and dispatch .github/workflows/ci.yml	Change detection enables serving-dspark for DSpark model changes. The job configures the runner, caches, timeout, pinned reference, and DSpark serving action invocation.

Priority: ⬇️ Low

Estimated code review effort: 3 (Moderate) | ~25 minutes

Sequence Diagram(s)

sequenceDiagram
  participant PullRequest
  participant detect-changes
  participant serving-dspark
  participant pypto-serving-tests
  participant task-submit
  PullRequest->>detect-changes: change models/deepseek_v4_flash_dspark/
  detect-changes->>serving-dspark: set run_dspark_serving_tests=true
  serving-dspark->>pypto-serving-tests: invoke deepseek-v4-dspark tests
  pypto-serving-tests->>task-submit: run accuracy test on 16 devices
  task-submit-->>pypto-serving-tests: return test result

Loading

Merge Risk: ⚪ Minimal · up to 24948

No confirmed merge-blocking risk remains in this CI-only change.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly identifies the main change: adding a CI serving accuracy guard for DSpark on 16 cards.
Description check	✅ Passed	The description directly matches the changeset. It covers the dedicated serving job, DSpark model configuration, 16-card accuracy testing, time limits, and cleanup.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 0…
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{A rabbit checks the model trail
DSpark tests hop through the rail
Sixteen devices start to run
Logs stay quiet, work gets done
CI marks the finish line done}

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

[ndleslx]

This draft depends on hw-native-sys/pypto-serving#213 landing first: the reusable action checks out pypto-serving/main, and that PR supplies tests/test_deepseek_dspark_accuracy.py and the DSpark serving integration.

The selected self-hosted runner needs 16 available cards and PYPTO_DSV4_DSPARK_MODEL_DIR set in its service environment to the DSpark W8A8 checkpoint directory.

Validation: all pre-commit hooks passed. Local checks parsed the YAML, validated embedded shell syntax, exercised DSpark versus unrelated change selection and missing-model handling, and captured the task-submit payload to verify the 16-card request, model path, PYTHONPATH, pytest target, and absence of Simpler timeout overrides.

The Palace 64/128 guard previously passed on the serving branch with pypto-lib 57e9d6a and default Simpler timeouts (2 passed in 500.94s). The new CI job has not been run end to end against this PR's pypto-lib base. As with the existing model-specific serving jobs, CI-only edits do not trigger the new device job.

Related: #1133