pd: a request larger than the page pool blocks the queue and spins a core #169

Problem

submit checks only max_context. If pages < max_context/page_size, a request that needs more pages than the pool can never be admitted, FIFO admission stops everything behind it, and the owner loop spins because the batcher never goes idle.

Evidence and reproduction

opt_prime/opt_prime/accel/pd/batching.py:177-179, pd_batching.py:363-365, paged.py:281. Ran: pages=2 with a 48-token request; neither it nor a small request behind it completed.

Environment

optimus-rm @253be20; Python 3.10, torch 2.10.0+cu128, transformers 5.1.0; RNGD+A100 host (CPU / one A100 for the checks).

Scope

PagedKV construction.

Done when

An undersized pool is rejected at construction; a test covers it.

Owner

@gspark-etri