[Leonccaa]

Draft, based on main. Replaces #71, which was stacked on the partial prefix-cache port (#70); this fix does not depend on it. GPU validation is planned for the next validation window.

Fixes #69

Change

The worker zeroes the pages in new_block_ids_to_zero in the step they are allocated, after the connector has submitted that step's pending stores. The connector fenced pending stores only against pages that reach it through scheduled requests' new blocks and offloaded groups' load destinations. A request parked for an async CPU load is not scheduled, and the QSA compressor-ring group is not offloaded, so a ring page it claimed could still be read by another request's pending store. The page was zeroed first and the CPU tier stored zeros under a valid prefix key.

build_connector_meta now flushes every pending store that reads a page in new_block_ids_to_zero. The extra flushes only happen when a page announced for zeroing still has a pending store.

The issue also affects 0.1.5; the change applies there unchanged.

Validation

CPU, in the ReVolta build image:

New tests in tests/v1/kv_connector/unit/offloading_connector/test_zeroing_fence.py (the file carries the minimal production cache layout it needs):

• test_zeroed_pages_are_fenced_against_pending_stores: a pending store on a page in new_block_ids_to_zero is flushed; other pages are not.
• test_parked_request_ring_page_is_fenced_before_zeroing: reproduces the issue with a real KVCacheManager in the production group layout (DCP2, 1,664-token pages). Request A offloads a drafter boundary state and moves past the page; request C is admitted with an async CPU load and receives that page as its compressor-ring page. Without the change jobs_to_flush is empty; with it the store is flushed.
• Both fail on the base commit and pass with the change.
• Regression over tests/v1/core, tests/v1/kv_offload and tests/v1/kv_connector/unit/offloading_connector: 823 passed, 253 failed, 26 errors, 6 skipped. The failures and errors are identical by node id to main (CUDA-only and pre-existing); the only additional passes are the two new tests, and no test that passes on main fails here.
• pre-commit (ruff, typos, mypy 3.12) passes on the changed files.

Not yet run on GPU.

Attribution

The defect was found in an independent review with an AI assistant; this change and its tests were written with an AI assistant (Claude Code). The submitting human reviews every changed line before merge. Commits are signed off under the DCO.

🤖 Generated with Claude Code

[Leonccaa]

GPU validation: the same fence change is part of the 0.2.0rc3 acceptance run on 4x V100 (offload readback, stress and long-context checks passed) and now running in production. Marking ready.