Prove draft HiCache mirrors target valid-tail pages

Draft HiCache already piggybacks target CP reservations, but the controller coverage only exercised page-aligned nodes. The new tests pin the intended contract for valid-tail nodes: target and draft reserve/load the same padded physical page while scheduler-visible load results keep only valid logical locs.

Constraint: Draft KV is a target mirror; it must not choose an independent valid-only host path.
Rejected: Re-enable draft partial-current reuse in this slice | current safe contract keeps EAGLE/NextN cache-hit draft on full materialization until same-layer padded visibility is proven.
Confidence: medium
Scope-risk: narrow
Directive: Keep draft and target host metadata coupled; do not add draft-only capacity or prefetch decisions.
Tested: local py_compile for test_hicache_controller_cp.py.
Tested: remote g0034 test_hicache_controller_cp.py: 59 passed, 3 warnings.
Not-tested: CUDA E2E runtime for this commit.
Co-authored-by: OmX <omx@oh-my-codex.dev>
This commit is contained in:
laoyao0822
2026-05-29 05:43:49 +08:00
co-authored by OmX
parent 0043037f78
commit 7e06eaebbf
2 changed files with 78 additions and 0 deletions
@@ -697,6 +697,9 @@ Current state:
`draft_host_indices` are present.
- Draft partial-current reuse/prefetch has been disabled or pushed to older
materialization paths in some cases to avoid EAGLE hangs.
- Existing draft HiCache controller tests cover page-aligned write/load/evict,
but they do not yet prove that a valid-tail node mirrors the target padded
physical span while exposing only the valid length.
Correction:
@@ -712,6 +715,15 @@ Tests:
`owned_positions` and padded physical length.
- EAGLE one-layer path does not prefetch or consume an unmasked padded tail.
Implemented C9 slice:
- Added target/draft write coverage for a valid-tail suffix. Target and draft
reserve the same padded local positions and transfer the same physical device
span while metadata keeps the valid `logical_len`.
- Added target/draft load coverage for a valid-tail host hit. Both pools load
the padded physical page, but `load_cp()` returns only valid logical locs to
the scheduler.
### C10. Disaggregated transfer has not been audited for valid/padded split
Current state: