[gspark-etri]

Background

A failed PD prefill producer could leave requests pending, keep /ready green and get the batcher quarantined on the next admission; producer stage durations were not exported. This lands that work from a draft stack, with two fixes found in review.

Changes

From feat/249-pd-stage-metrics (#239, #247, #250):

• opt_prime/opt_prime/accel/pd/pd_batching.py: SubprocessPrefillWorker settles every pending future on child exit or a malformed reply, rejects new work after a terminal failure, cleans up the child and shared memory on failed startup, exposes ready; PrefillWorkerPool.ready.
• opt_prime/opt_prime/accel/serve/batched.py: BatchedService.ready, admission returns 503 retryable while the producer is unready, airm_pd_producer_seconds_total{stage=prefill|transfer}.
• opt_prime/opt_prime/accel/serve/llmd.py: only the /ready hunk (from exp/240-gpu-live-pd...fix/246-pd-producer-readiness); the rest of the file is untouched so the SIGPIPE fix in the serving PR is not reverted.
• Tests: test_subprocess_prefill_failure.py, test_pd_producer_readiness.py, test_pd_stage_metrics.py (new).

Fixes on top (test first):

• F1: SubprocessPrefillWorker.close() now closes stdin and waits for the child to exit, then terminate(), then kill(). prefill_server exits its for line in sys.stdin loop on EOF; SIGTERM first skipped the SDK cleanup that releases the RNGD device. Test test_close_lets_child_exit_cleanly_on_stdin_eof. The existing test that a child ignoring SIGTERM is still killed is kept and passes.
• F2: ProcessPrefillWorker.ready (self._futures is not None and self._proc.is_alive()), so the readiness gate sees a dead child. Test test_dead_process_worker_is_unready_and_rejects_admission: child killed, /ready 503, completion 503 retryable, service not poisoned.

Verification

• pytest -q tests/test_subprocess_prefill_failure.py tests/test_pd_producer_readiness.py tests/test_pd_stage_metrics.py: 16 passed.
• Before the fixes: F1 test failed (returncode -15 != 0, no clean-exit marker); F2 test failed (/ready returned 200 after the child was killed).
• pytest -q tests/test_pd_batching_disagg.py (existing tests of both worker classes): 12 passed.
• Full suite: CI (cpu-tests).

Compatibility and limits

• close() can now take up to 2 * timeout (default 5 s each) before SIGKILL for a child that ignores stdin EOF and SIGTERM.
• A failed producer is not replaced or removed from a pool automatically; replacing it is explicit.
• The duration metrics are worker-reported seconds for admitted handoffs, not HTTP latency.
• This PR and the serving PR both edit llmd.py, in separate hunks.

Related Issue

Closes #238
Closes #246
Closes #249

Supersedes #239, #247, #250.

🤖 Generated with Claude Code