[yifjiang]

Summary

Use retained GPU inventory as a target baseline after native eviction without granting execution credit. Preserve authoritative attachment and topology precedence; LLM wrappers follow in D13.

Review slice D12 of #16748; part of #16747 and #16746. This draft targets codex/kv-reuse-d11-bounded-lookups for a complete incremental review diff of 499 additions plus deletions, including implementation and tests.

The ZMQ integration fixture waits for its listener's subscription before sending the original one-shot GPU and CPU events. Payloads, sequence numbers, the five-second convergence check and tier assertions remain unchanged. No diagnostic listener delay enters the source.

Reuse remains experimental and default off. Supported activation requires the companion KVCR/native slices and declared deployment profile; retained inventory is advisory and native operations retain ownership until quiescence.

Validation

A controlled 300ms late subscriber makes the original fixture genuinely fail and the subscription barrier pass 22 case executions. The clean D12 fixture-only source passed 177 integration executions, allocated formatting and all-target Clippy. The changed-file hook scope also passed; the inherited D02 HTTP source has a separate exact-tree hook receipt.

These controls qualify subscription readiness and preserve the original production listener. They do not establish the original hosted timeout's cause. The inherited HTTP fix separately passes four focused regressions, all 60 service cases, 150 unchanged HTTP executions and strict all-target Clippy. Current-head hosted CI remains required; native transport and hardware results retain their recorded producer scope.

[devin-ai-integration]

❌ Dynamo PR CI failed — run 41803083245 (attempt 1) on 7c31512de2

Gate checks: ✅ backend-status-check · ✅ deploy-status-check · ❌ dynamo-status-check

Other	Jobs
dynamo-runtime	❌ 2 ✅ 5

Failure details

2 jobs failed (amd64 and arm64, same 2 tests): test_shared_router_settings_fixture_flows_to_binding_kwargs fails with ValueError: unknown router setting(s) for the decode-to-prefill transfer settings, and test_no_registered_env_var_is_read_from_os_environ sees reject_media_status_code 0 instead of 415. No cancelled jobs.

❌ dynamo-runtime / test / sequential cuda13.0, amd64: ValueError unknown router setting(s); reject_media_status_code 0 != 415

Job: dynamo-runtime / test / sequential cuda13.0, amd64 · Failed step: Run CPU-only tests (parallelized) · Logs: gh run view --job 126496502973 -R ai-dynamo/dynamo --log-failed

FAILED components/src/dynamo/common/tests/configuration/test_kv_router_args.py::test_shared_router_settings_fixture_flows_to_binding_kwargs - ValueError: unknown router setting(s): ['enable_decode_to_prefill_gpu_reuse', 'enable_decode_to_prefill_reuse', 'minimum_estimated_ttft_saving_ms', 'minimum_transfer_prefix_blocks', 'transfer_source_estimated_block_us', 'transfer_source_gpu_estimated_block_us', 'transfer_source_lookup_budget_ms', 'transfer_source_prefill_token_us']
  kv_router_args.py:296 in kv_router_kwargs -> :248 apply_router_config -> :234 _router_settings
FAILED components/src/dynamo/frontend/tests/test_frontend_parser_builder.py::test_no_registered_env_var_is_read_from_os_environ - AssertionError: 0
E     Differing items:
E     {'reject_media_status_code': 0} != {'reject_media_status_code': 415}

Two Python unit tests fail (after 3 auto-retries each): _router_settings rejects 8 decode-to-prefill transfer/reuse settings as unknown router settings, and the frontend config built without env vars has reject_media_status_code = 0 instead of the default 415. Deterministic test failures, not infrastructure.

❌ dynamo-runtime / test / sequential cuda13.0, arm64: ValueError unknown router setting(s); reject_media_status_code 0 != 415

Job: dynamo-runtime / test / sequential cuda13.0, arm64 · Failed step: Run CPU-only tests (parallelized) · Logs: gh run view --job 126496503205 -R ai-dynamo/dynamo --log-failed

FAILED components/src/dynamo/common/tests/configuration/test_kv_router_args.py::test_shared_router_settings_fixture_flows_to_binding_kwargs - ValueError: unknown router setting(s): ['enable_decode_to_prefill_gpu_reuse', 'enable_decode_to_prefill_reuse', 'minimum_estimated_ttft_saving_ms', 'minimum_transfer_prefix_blocks', 'transfer_source_estimated_block_us', 'transfer_source_gpu_estimated_block_us', 'transfer_source_lookup_budget_ms', 'transfer_source_prefill_token_us']
  kv_router_args.py:296 in kv_router_kwargs -> :248 apply_router_config -> :234 _router_settings
FAILED components/src/dynamo/frontend/tests/test_frontend_parser_builder.py::test_no_registered_env_var_is_read_from_os_environ - AssertionError: 0
E     Differing items:
E     {'reject_media_status_code': 0} != {'reject_media_status_code': 415}

Two Python unit tests fail (after 3 auto-retries each): _router_settings rejects 8 decode-to-prefill transfer/reuse settings as unknown router settings, and the frontend config built without env vars has reject_media_status_code = 0 instead of the default 415. Deterministic test failures, not infrastructure. Same failures as the amd64 job.

For agents

{"pr": 18030, "run_id": 41803083245, "run_attempt": 1, "head_sha": "7c31512de2effbaca1b809178d68b2ddb8a54b1d", "failures": [{"job": "dynamo-runtime / test / sequential cuda13.0, amd64", "job_id": 126496502973, "failed_step": "Run CPU-only tests (parallelized)", "signature": "ValueError: unknown router setting(s) [decode-to-prefill transfer settings]; AssertionError reject_media_status_code 0 != 415", "tests": ["components/src/dynamo/common/tests/configuration/test_kv_router_args.py::test_shared_router_settings_fixture_flows_to_binding_kwargs", "components/src/dynamo/frontend/tests/test_frontend_parser_builder.py::test_no_registered_env_var_is_read_from_os_environ"], "log_cmd": "gh run view --job 126496502973 -R ai-dynamo/dynamo --log-failed"}, {"job": "dynamo-runtime / test / sequential cuda13.0, arm64", "job_id": 126496503205, "failed_step": "Run CPU-only tests (parallelized)", "signature": "ValueError: unknown router setting(s) [decode-to-prefill transfer settings]; AssertionError reject_media_status_code 0 != 415", "tests": ["components/src/dynamo/common/tests/configuration/test_kv_router_args.py::test_shared_router_settings_fixture_flows_to_binding_kwargs", "components/src/dynamo/frontend/tests/test_frontend_parser_builder.py::test_no_registered_env_var_is_read_from_os_environ"], "log_cmd": "gh run view --job 126496503205 -R ai-dynamo/dynamo --log-failed"}]}

_{Posted automatically by Devin for run 41803083245. Updated on every full-CI run of this PR.}

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[yifjiang]

/ok to test c76acf8

[yifjiang]

/ok to test 7cbfc3c