[HiSparse] Switch to host reads once a request fills its admission window

[MatthewBonanni]

Stacked on #59494. Optional: #59494 alone fixes the livelock; this avoids the preemptions behind it.

Purpose

Admission caps each request's resident pages at the in-flight window (HiSparseResidentSpec.max_admission_blocks_per_request, derived from max_in_flight_tokens), on the assumption that older pages move to host. At runtime, though, a request keeps its resident pages pinned until free GPU blocks fall below a fixed watermark (a tenth of the pool). A chunked prefill whose next chunk needs more than that outgrows the pool it was admitted into and is preempted, even though its written-back pages could have been released. With #59494 it resumes from host, but the preemption is still wasted work.

Changes

HiSparseCoordinator.update_residency also switches a request to host reads (asks for a hot region, then releases its written-back pages) once its resident pages fill the window admission reserved for it. Runtime then enforces the same assumption admission makes, with no second estimate. HiSparseResidentManager exposes that window as a read-only max_admission_blocks_per_request property; nothing else in vllm/v1/core changes.

Released pages stay readable until the pool reuses them, so a long request that switches early only pays for its hot region.

Test Plan

• New test_hisparse_prefill_reads_host_once_it_fills_admission_window[4-True, 3-False]: a prefill that fills its admission window requests a hot region while free blocks are still above the watermark; one below it does not.
• Same H100 engine repro as [Bugfix][HiSparse] Fix a chunked-prefill preemption livelock #59494 (shrunk DeepSeek-V3.2, indexer + 8 resident groups sharing the GPU pool), tests/v1/e2e/general/test_hisparse.py, and HiSparse unit tests (-k hisparse).

Test Result

Repro	main	#59494	this PR (on #59494)
1×120k prompt, 4000 blocks, 8192-token chunks	livelock	35 steps, 1 preemption	32 steps, 0 preemptions
1×120k prompt, 12000 blocks	livelock	26 steps, 1 preemption	23 steps, 0 preemptions
2×60k prompts, 4000 blocks	livelock	40 steps, 2 preemptions	34 steps, 0 preemptions
3×40k prompts, 6000 blocks	35 steps, 2 preemptions	29 steps, 2 preemptions	24 steps, 0 preemptions
1×120k, 1024-token chunks	126 steps	126 steps	126 steps

• The 4-True case fails on [Bugfix][HiSparse] Fix a chunked-prefill preemption livelock #59494 alone. test_prefix_caching.py: 168 passed (CPU).
• H100: test_hisparse.py 4 passed; HiSparse unit tests 165 passed (plus [Bugfix][HiSparse] Fix a chunked-prefill preemption livelock #59494's updated assertion).

AI assistance (Claude Code) was used for the investigation, implementation, tests, and this description.

🤖 Generated with Claude Code

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

[MatthewBonanni]

Folded into #59494.