Remove inproc_rpc - #768
Open
melody-ren wants to merge 12 commits into
Open
Conversation
inproc_rpc was a value of CUDAQ_QEC_REALTIME_MODE that routed realtime
enqueue/get/reset through an in-process shared-ring RPC session
(qec_realtime_session + rpc_producer) instead of the direct-call path.
Remove it entirely; the direct-call path becomes the sole in-process path.
The external_server (two-process decoding server) path is unaffected -- it
never keyed on inproc_rpc.
- realtime_decoding.{cpp,h}: remove the session setup, the three
g_realtime_session RPC branches, get_realtime_session(), and the
inproc-only cudaDeviceMapHost flag. The surviving direct-call path
(validation, --save_syndrome capture, device pinning, decode) is unchanged.
- Delete qec_realtime_session.{cpp,h} and rpc_producer.{cpp,h} and drop them
from the realtime-decoding library; nothing else in production used them
(the decoding servers only mirror their memory layout).
- Disable the inproc-only tests (test_pymatching_realtime,
test_realtime_qldpc_graph_decoding, the surface_code-1 inproc-rpc ctest,
and the sc4-yaml REQUIRE_HOST_MODE assertion) pending removal and re-point
in follow-up commits.
Signed-off-by: Melody Ren <melodyr@nvidia.com>
Follow-up to removing inproc_rpc: delete the tests that existed only to exercise the in-process RPC session, and move the app-example tests off the now-ignored inproc_rpc env onto the surviving decode paths. - Delete the inproc-only unit tests test_pymatching_realtime and test_realtime_qldpc_graph_decoding (+ qldpc_config_loader), and the surface_code-1 inproc-rpc CTest. - Re-point surface_code-4-yaml / surface_code-1 to the direct-call path (drop CUDAQ_QEC_REALTIME_MODE, now unread). External-server / UDP coverage is unchanged -- the cqr variants still route through the two-process server. - Remove the REQUIRE_HOST_MODE assertions (they only proved the inproc session was used). - Disable the relay-BP DEVICE-mode cases (nv-qldpc --use-relay-bp, trio-relay-d3) that drove the removed device-graph scheduler; re-point them to the external server (HOST_CALL over UDP) as a follow-up. Signed-off-by: Melody Ren <melodyr@nvidia.com>
- realtime_relay_bp.rst: excise the CI unit test and Surface Code Test sections (both drove the removed inproc_rpc device-graph path) and the CI-test build scaffolding (config list, hardware table, 'CI unit test only' build). Keep the emulated / FPGA end-to-end content, which drives the same device-graph scheduler through the Hololink bridge. - Re-anchor comments that named the deleted qec_realtime_session: the decoding-server ABI/layout notes (GpuRoceTransceiver, decoding_server) point at cudaq-realtime or describe the mechanism directly; the Hololink bridge and the realtime-decoding CMake drop the qec_realtime_session references. Signed-off-by: Melody Ren <melodyr@nvidia.com>
Conflict resolutions, all in favor of upstream's post-NVIDIA#682/NVIDIA#750 realtime structure with the inproc_rpc path kept out: - qec_realtime_session.cpp: deleted (upstream's touch was to removed code). - GpuRoceTransceiver.cpp: deleted; NVIDIA#754 renamed it to DeviceGraphTransceiver.cpp. The stale qec_realtime_session comment reference the branch scrubbed moved to DeviceGraphRingConsumer.cpp and is scrubbed there instead. - realtime_decoding.cpp: reset_decoder takes NVIDIA#698's pin_decode_device(), dropping the rpc_producer dispatch branch. - decoding_server.cpp: upstream's per-ring rewrite; its comments no longer reference qec_realtime_session, so the branch's scrub is subsumed. - app_examples/CMakeLists.txt: upstream's add_surface_code_4_yaml_test signature (onnx_path folded into ARGN) with an empty test environment. - surface_code-4-yaml-test.sh: keep the num_logical echo, drop the CUDAQ_QEC_REALTIME_MODE echo. - realtime_relay_bp.rst: single Building subsection renamed Hololink -> HSB, target list on the NVIDIA#754 names, CI-unit-test section dropped (the test is removed) and the Surface Code Test section kept. Upstream added no new inproc_rpc surface: DeviceGraphTransceiver, DeviceGraphRingConsumer, surface_code-5-per-decoder-rings and realtime_decoding_demo are all inproc-free. Signed-off-by: Melody Ren <melodyr@nvidia.com>
The inproc_rpc removal parked these behind if(FALSE) on the theory that they needed the device-graph scheduler. They did not: --use-relay-bp only sets nv-qldpc parameters when the decoder config is generated, so the cases need the plugin and nothing else. Re-enable them on the direct-call path, on the same gate the dual-parse cases already use, and fold that repeated gate into _SC4_HAVE_NV_QLDPC. What the inproc_rpc cases did provide was relay-BP over a *dispatched* path. Restore that against the external decoding server, which since NVIDIA#682 gives each decoder its own ring: one all-nv-qldpc trio, and the pymatching + trt + nv-qldpc trio when the trt plugin and python-onnx are present. Both assert server-side decode counts, and both get a 600s timeout because relay BP iterates. Signed-off-by: Melody Ren <melodyr@nvidia.com>
realtime_relay_bp.rst still compared the surface-code test against the CI unit test section, and pointed at it for the build, but that section went away with the test it documented. Say what the app example does on its own terms and point the build note at the bridge. The surface_code-4 comment naming inproc_rpc described what the external relay-BP cases replaced rather than what they are. Signed-off-by: Melody Ren <melodyr@nvidia.com>
The new surface_code-4 external-server cases (cqr-relay-bp, cqr-trio-relay) were gated on _SC4_HAVE_NV_QLDPC, i.e. on the decoder plugin alone. That is the right gate for the direct-call cases -- --use-relay-bp only sets nv-qldpc parameters when the config is generated -- but not for these two: the server constructs the decoder through DecodingSession::create(), which probes supports_graph_dispatch() and captures a CUDA device graph even on the udp transport. Without decoding_server linked against the proprietary cudevice archive that lookup fails and the server never reaches READY: nv_qldpc_decoder::capture_decode_graph: cudaqx_qec_realtime_dispatch_register_decoder_state not found via dlsym So in any build with the plugin but no archive -- which CMake already warns about at decoding_server link time -- the tests registered and hard-failed. Both plugin flavors in the proprietary tree behave identically here, so this is not a host-vs-graph build question. Add _SC4_NV_QLDPC_IN_SERVER (plugin AND device-graph availability AND the proprietary target, mirroring tools/decoding-server) and hang the two external cases off it. The direct-call cases keep the plugin-only gate. Also warn when either nv-qldpc gate fails, per the convention the trt block right below already follows -- a silent skip here is precisely what once masked a dead test check for a whole release cycle. Verified in all three configurations: plugin+archive registers and passes (47/47 surface_code-4 cases, including cqr-trio-relay); plugin without the archive skips the external cases with a named warning while the direct-call cases still run; no plugin skips all of them. Signed-off-by: Ben Howe <bhowe@nvidia.com>
No conflicts. The nine commits since the last merge touch the realtime path again -- NVIDIA#769 serves HOST_CALL inline and retires the worker threads, queues, and CqrTransceiver; NVIDIA#759 adds the dynamic-DEM infrastructure that reshapes create_realtime_decoder around chunk expansion -- but none of it reaches the inproc_rpc path this branch removes: no reference to inproc_rpc, qec_realtime_session, or rpc_producer comes back, and none of the deleted sources reappear. NVIDIA#778 moves the CUDA-Q pin to 28f195f4, so verification needs a toolchain rebuilt at that commit. Signed-off-by: Melody Ren <melodyr@nvidia.com>
…daqx into melodyr/goodbye-inproc-rpc
surface_code-4-yaml-mixed-dispatch already runs nv-qldpc RelayBP through the decoding server and asserts on its per-decoder logical_errors line, so these five registrations duplicated coverage that already exists. They were also unsound as written. Multi-type mode replaces the aggregate logical-error ceiling with a per-decoder one, but that check compares a single patch against MAX_NON_ZERO, which is the per-logical bound scaled by the patch count -- three times too loose with three patches. At the shot counts used here an nv-qldpc entry that decoded nothing would have stayed under it, and REQUIRE_SERVER_DECODE_COUNTS only proves the decoder was reached. Restoring this coverage belongs with a fix to that ceiling, not in the inproc_rpc removal. _SC4_NV_QLDPC_IN_SERVER and its warning branch go with them; _SC4_HAVE_NV_QLDPC stays for the dual-parse cases. Signed-off-by: Melody Ren <melodyr@nvidia.com>
The library's two sources, realtime_decoding.cpp and config.cpp, name no host-dispatch symbol; the link existed for qec_realtime_session.cpp, which this branch removes. cudaq_host_dispatcher_loop still has a caller in realtime_pipeline.cu, but that belongs to the separate cudaq-realtime-pipeline target with its own link, so only this target's lookup and link go. Signed-off-by: Melody Ren <melodyr@nvidia.com>
melody-ren
marked this pull request as ready for review
August 15, 2026 06:09
Collaborator
Author
|
/ok to test 8d82248 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Removes
inproc_rpc, the in-process RPC session that was an intermediary stepin the decoding-server work. The external decoding server (HOST_CALL over UDP)
and the direct-call path are the surviving decode paths.
qec_realtime_sessionandrpc_producer, and the tests that existedonly to exercise them (
test_pymatching_realtime,test_realtime_qldpc_graph_decoding+qldpc_config_loader, thesurface_code-1inproc CTest).CUDAQ_QEC_REALTIME_MODE(now unread) and theREQUIRE_HOST_MODEassertions, which only proved the inproc session was in use.
realtime_relay_bp.rstsections that documented the removedpath; re-anchor comments naming the deleted session.
cudaq-qec-realtime-decodingagainstlibcudaq-realtime-host-dispatch; that link existed forqec_realtime_session.cpp, and neither remaining source names ahost-dispatch symbol.
relay-BP coverage
Every
surface_code-4-yamlcase previously ran underCUDAQ_QEC_REALTIME_MODE=inproc_rpc. pymatching and trt fall through to thedirect-call path; the nv-qldpc/relay-BP cases were parked behind
if(FALSE).This PR deletes them rather than restoring them.
surface_code-4-yaml-mixed-dispatchalready runs nv-qldpc RelayBP throughdecoding_serverand asserts on its per-decoderlogical_errorsline, underthe same device-graph + proprietary-archive gate those cases would have
required — so restoring them adds no configuration that is not already
covered.
Testing
42/42
app_examples.surface_code-4-yamltests passwith the nv-qldpc plugin, proprietary cudevice archive, and python-onnx
configured. This covers the surviving direct-call and external decoding-server
paths after removing
inproc_rpc, includingapp_examples.surface_code-4-yaml-mixed-dispatch, where nv-qldpc RelayBP runson a device-graph ring alongside a host-dispatched decoder.