[fzyzcjy]

---

CI States

Latest PR Test (Base): ❌ Run #38503948215
Latest PR Test (Extra): ❌ Run #38503948193
Latest PR Test (AMD ROCm 10): ❌ Run #38503948175

[fzyzcjy]

/rerun-test test/registered/rl/test_weight_version_spans.py

[github-actions]

⚠️ Rebase Required Before Re-run

A major update has landed on main. Your PR is diverged relative to required base commit 60878c3fba50.

Re-run was not dispatched. What to do:

• Rebase your branch onto the latest main and push again
• Follow issue #21065 for context
• CI-fix PRs may request the bypass-maintenance label to skip this check

[fzyzcjy]

AI-assisted validation report (Codex), requested by the PR author. Not ready to merge: the exact head has failing new unit tests.

• Head: 653fbddb34e5bc4f989f7b1330e686619f91564f; scope: sglang-miles...head (the intended base). No source/test changes were made.
• Remote: 1 NVIDIA RTX 4090, lmsysorg/sglang:latest, Python 3.12.3, PyTorch 2.13.0+cu130, pytest 9.1.1. Imports resolve to the synced exact-head source.
• Pytest summary: 422 passed, 38 skipped, 4 failed, plus 127 passing subtests. The four failures comprise two failed methods and two failed subtests of a third method; JUnit therefore records 3 failing parent test cases, not four independent methods.

Test file (under test/)	Passed parents	Failed parents	Skipped
registered/unit/hardware_backend/mlx/test_attention_patching.py	0	0	38
registered/unit/managers/test_batch_result_processor_hidden_states.py	9	0	0
registered/unit/managers/test_batch_result_processor_mamba_boundary.py	1	0	0
registered/unit/managers/test_batch_result_processor_spec_grammar.py	7	0	0
registered/unit/managers/test_msgpack_ipc_roundtrip.py	10	0	0
registered/unit/managers/test_multi_tokenizer_mixin.py	9	0	0
registered/unit/managers/test_output_streamer_customized_info.py	15	0	0
registered/unit/managers/test_scheduler_weight_version_tracking.py	24	3	0
registered/unit/mem_cache/test_kv_weight_version_tracker.py	24	0	0
registered/unit/model_executor/test_model_runner_kv_weight_versions.py	3	0	0
registered/unit/server_args/test_server_args.py	282	0	0
registered/unit/utils/test_weight_versions.py	37	0	0

Failure: test_scheduler_weight_version_tracking.py constructs Scheduler with __new__ but omits tree_cache. Scheduler.process_batch_result() unconditionally calls self.tree_cache.flush_pending_backups() at scheduler.py:4818, raising AttributeError in:

• test_pending_result_uses_version_captured_when_batch_ran
• test_result_copy_finishes_before_slot_indices_are_consumed
• test_spec_decode_does_not_restamp_restored_prefill_slots (overlap=False and True)

Suggested correction: initialize the fixture's unrelated cache dependency (tree_cache with flush_pending_backups), retaining all existing KV-version assertions. This has not been applied or validated yet.

Reproduction on the exact head in the environment above:

python -m pytest -q test/registered/unit/hardware_backend/mlx/test_attention_patching.py test/registered/unit/managers/test_batch_result_processor_hidden_states.py test/registered/unit/managers/test_batch_result_processor_mamba_boundary.py test/registered/unit/managers/test_batch_result_processor_spec_grammar.py test/registered/unit/managers/test_msgpack_ipc_roundtrip.py test/registered/unit/managers/test_multi_tokenizer_mixin.py test/registered/unit/managers/test_output_streamer_customized_info.py test/registered/unit/managers/test_scheduler_weight_version_tracking.py test/registered/unit/mem_cache/test_kv_weight_version_tracker.py test/registered/unit/model_executor/test_model_runner_kv_weight_versions.py test/registered/unit/server_args/test_server_args.py test/registered/unit/utils/test_weight_versions.py --junitxml=/scratch/pr44164-unit.xml

Limitations:

• All 38 MLX cases skipped: no Apple/MLX execution coverage.
• The modified two-GPU RL file test/registered/rl/test_weight_version_spans.py is not validated yet; remote validation is being prepared.
• Lint passed, but normal CI did not execute the tests. The targeted rerun was rejected by the main maintenance gate; this PR deliberately targets sglang-miles. No bypass or base change was applied.
• Full local evidence: /Volumes/MyExternal/domains/agent-default/artifacts/sglang/2026-10-14-pr44164-validation/ (unit.log, unit.xml, unit-xml-env.log, sync.log, environment.log); these are local paths, not publicly downloadable CI artifacts.

[fzyzcjy]

AI-assisted validation follow-up (Codex), requested by the PR author.

• Head unchanged: 653fbddb34e5bc4f989f7b1330e686619f91564f; no source, test, threshold, or server argument modifications.
• RL attempt: test/registered/rl/test_weight_version_spans.py, two NVIDIA B200 GPUs, lmsysorg/sglang:latest, original TP=2 / DP=2 / EAGLE configuration. H100/H200 capacity was unavailable; the earlier 4090 was released before this allocation.
• Result: 24 setup errors, 0 test bodies executed. The server failed to download lmsys/sglang-ci-dsv3-test/resolve/main/config.json with HTTP 401 Unauthorized. The model was absent from the shared cache. This is an environment/access blocker, not evidence of an RL assertion failure or successful feature coverage.
• Reproduction: HF_HOME=/scratch/pr44164-hf timeout 1800 python -m pytest -q -s test/registered/rl/test_weight_version_spans.py --junitxml=/scratch/pr44164-rl.xml.
• Artifacts: /Volumes/MyExternal/domains/agent-default/artifacts/sglang/2026-10-14-pr44164-validation/rl.log, rl-xml-env.log, rl-environment.log, and rl-sync.log (local retained artifacts, not public CI downloads).
• Merge recommendation remains no: the independently reproduced unit failures still need correction; RL coverage remains missing; the main maintenance gate also blocks CI for this deliberately sglang-miles-based PR. No bypass label or base change was applied.

[fzyzcjy]

AI-assisted follow-up (Codex), requested by the PR author.

• Amended and pushed: 99d65b886cdbe7675705082d5007c6d900e08cf2; the PR still contains one commit onto sglang-miles.
• Incremental fix: only test/registered/unit/managers/test_scheduler_weight_version_tracking.py, +2/-1. The fixture now supplies tree_cache.flush_pending_backups, sharing that cache with SchedulerBatchResultProcessor. No production changes, assertion removals, threshold changes, skip additions, or altered scenarios.
• Remote targeted regression: the three previously failing parent methods now pass: 3 passed, 2 subtests passed.
• Remote final run: all original 12 changed unit-test files rerun at the new head: 424 passed, 38 skipped, 129 subtests passed, 0 failures. All 38 skips require MLX; no Apple backend coverage is claimed.
• Environment: one RTX 4090, lmsysorg/sglang:latest, exact committed source synced using tom-remote; the resource has been unbound and released. The changed test's remote and local SHA-256 is a0b8d442378d71a78280f5251566f0e021b175074cb23ca412d72aaa21280f73.
• Lint: all applicable hooks passed for the modified test file. Full-repository pre-commit run --all-files was attempted but cannot be reported green: local environment lacks PyYAML/Rust tooling. The targeted lint rerun used a temporary uv environment with PyYAML. No lint changes were produced.
• Remaining limits: RL was not rerun for this fixture-only amendment. Its prior two-GPU attempt was blocked by HTTP 401 fetching the required model, with zero test bodies executed. The separate CI maintenance gate remains unresolved; this is not an all-CI-green or full merge-gate PASS claim.
• Retained evidence: /Volumes/MyExternal/domains/agent-default/artifacts/sglang/2026-10-14-pr44164-validation/fixture-targeted.log, fixture-full.log, fixture-xml-status.log, fixture-sync.log, and fixture-pre-commit*.log. These local artifacts contain exact commands and test output; they are not public CI downloads. Remote git status hit the image's ownership check; source identity is supported by clean local commit + sync record + matching test hash, not a claim that the image's Git metadata equals the PR head.

Reproduce the focused regression at the new head:

python -m pytest -q \
  test/registered/unit/managers/test_scheduler_weight_version_tracking.py::TestSchedulerBatchWeightVersion::test_pending_result_uses_version_captured_when_batch_ran \
  test/registered/unit/managers/test_scheduler_weight_version_tracking.py::TestSchedulerBatchWeightVersion::test_result_copy_finishes_before_slot_indices_are_consumed \
  test/registered/unit/managers/test_scheduler_weight_version_tracking.py::TestSchedulerBatchWeightVersion::test_spec_decode_does_not_restamp_restored_prefill_slots