-
[Refactor] Group prefetch params into PrefetchRequestSpec (#4121)
frostbyte_neo 发布于
2026-07-22 19:55:48 +00:00 - [Refactor] Group prefetch params into PrefetchRequestSpec
Introduce a frozen PrefetchRequestSpec dataclass in distributed/api.py
(alongside PrefetchHandle) bundling the six L2-fetch inputs: keys,
layout_desc, extra_count, policy, attn_desc, mode.- PrefetchController.submit_prefetch_request(spec): replaces the six
flat params and removes the fragile positional submission tuple; the
field list now lives in one place. - StorageManager.submit_prefetch_task(spec, external_request_id='',
skip_l2=False): takes the shared spec plus the two StorageManager-only
orchestration knobs as explicit args (kept out of the spec so the
controller's contract has no settable-but-ignored fields). - Update the six MP-module callers and all affected tests.
No behavior change. mypy + non-CUDA mock tests pass; CUDA-gated prefetch
tests require GPU CI.Signed-off-by: aeon-x talexcao@gmail.com
下载附件