[valarLip]

Deferred speculative outputs retain pinned PyTorch storage through NumPy views, and their D2H destinations are recreated each step. Return independently owned NumPy arrays while reusing pinned destinations and completion events. Both prefill and verification now copy the two speculative status counts in one D2H operation.

• D2HQueue owns FIFO ordering, slot reuse and completion. Capacity comes from max_num_seqs (with the existing two-request block-drafter warmup floor); changing batch size reuses storage and copies only the active region. No per-transfer handle is allocated.
• SpecStatusQueue specializes only the output snapshot, copying each row once in scheduler order. The runner supplies contiguous [bonus, rejected] counts from the final post-broadcast verdict or a persistent prefill zero buffer, without a GPU packing kernel.
• ScheduledBatchOutput detaches non-owning NumPy views and retains already-owned arrays. Received outputs survive later transfers and cleanup without retaining pinned Torch storage.

Validation on main 65a199ea8, in rocm/atom-dev:nightly_202610151507, using only physical GPU 4:

• 433 passed, no failures or skips: transfer/storage, deferred speculative status, prefill postprocessing, DSpark, scheduler and pipeline-parallel tests, including real asynchronous copies.
• Independent CPU tests with GPU visibility disabled and AITER imports prohibited: 6 passed, two GPU-only tests skipped as intended.
• Profiling 64 → 63 → 64 → 1 → 128 verifies no new destination allocation after warmup and one status copy for both prefill and verification. Black, Ruff and diff checks pass.

The naming cleanup changes no runtime operations (checked by normalizing identifiers and comparing ASTs). Regression tests are limited to three new modules: transfer invariants with CPU doubles and real GPU copies, runner wiring/capacity, and output-storage ownership.

A same-process transfer microbenchmark against the earlier PR revision (2fe751209) used 31 randomized A/B pairs, 200 send/receive operations per block. Prefill paired median time decreased 34.2% at BS64 (95% bootstrap interval: −37.2% to −24.3%) and 19.4% with changing batch sizes (−24.4% to −15.5%). Draft and verified-status differences were inconclusive; their intervals include zero. These measure synchronous D2H send/receive round trips, not end-to-end inference throughput or latency, and compare against the previous PR implementation rather than main.

[github-actions]

🏷️ CI Guide

Runs automatically on every eligible PR before approval:

• ✅ Pre Checkin: Black, Ruff, catalog schema validation, non-GPU unit tests

Heavy model tests:

• ✅ Run after the PR is approved and Pre Checkin passes
• ✅ Run immediately when an approval review is submitted
• ✅ Can be requested before approval with labels

Label	Tests
ci:full	Run all heavy PR model tests: native ATOM, vLLM, and SGLang
ci:atom	Run native ATOM model accuracy tests
ci:vllm	Run ATOM vLLM OOT model accuracy tests
ci:sglang	Run ATOM SGLang model accuracy tests

Heavy jobs are skipped when the PR is not approved and no matching ci:* label is present.
Add labels via the sidebar or gh pr edit 2670 --add-label <label>