Repository navigation
test(rccl): add host-only unit tests for src/transport/net.cc - #13548
Merged
Merged
Conversation
1 task done
✅ All Policy Checks Passed
📖 Need help? See the Policy FAQ for details on every check and how to fix failures. |
![amd-argus[bot]](https://avatars.githubusercontent.com/in/4740020?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
ARGUS
Findings
5 more finding(s), not posted inline
projects/rccl/test/host/net-test.cc:203(Low) No test uses a known vendor with a non-matching device id, so the device-prefix check at net.cc:300 is unpinnedprojects/rccl/test/test_categories_micro_net.yaml:3(Low) test_categories_micro_net.yaml header claims setup/connect/proxy-progress/buffer-pool coverage the suites do not haveprojects/rccl/test/host/net-test.cc:148(Low) ReadIbNicRate_SeveralPorts puts the max rate in the last port, so a first- or last-parsable regression can pass depending on readdir orderprojects/rccl/test/host/net-test.cc:217(Low) Detect_MissingPciIds only covers both id files missing, so a||to&&change at net.cc:294 still passesprojects/rccl/test/host/net-test.cc:21(Nit) net-test.cc uses std::function, std::vector, std::pair and std::call_once without including <functional>, <vector>, <utility>, <mutex>
Bottom line: Looks approval ready.
Comment @argus review to re-run, or @argus re-review after pushing.
dhanavikram
changed the title
This was referenced Oct 15, 2026
Contributor
There was a problem hiding this comment.
ARGUS
This PR changes 5,560 lines across 22 files. A change this size is hard to review
well, for a person and for this bot alike: reviewers lose the thread, defects hide in the
volume, and the feedback arrives too late to be cheap to act on. Splitting it into a
stacked pull request
gives each piece a reviewable diff and keeps the history. This review covers the whole change
regardless.
Bottom line: LGTM.
Comment @argus review to re-run, or @argus re-review after pushing.
aduforest
approved these changes
Oct 16, 2026
pvallem
approved these changes
Oct 16, 2026
1 task done
dhanavikram
force-pushed
the
users/dhanavikram/net-helpers-host-tests
branch
from
October 16, 2026 16:28
8f05b1b to
8614108
Compare
Contributor
There was a problem hiding this comment.
ARGUS
This PR changes 5,569 lines across 23 files. A change this size is hard to review
well, for a person and for this bot alike: reviewers lose the thread, defects hide in the
volume, and the feedback arrives too late to be cheap to act on. Splitting it into a
stacked pull request
gives each piece a reviewable diff and keeps the history. This review covers the whole change
regardless.
Bottom line: LGTM.
Comment @argus review to re-run, or @argus re-review after pushing.
dhanavikram
force-pushed
the
users/dhanavikram/net-helpers-host-tests
branch
from
October 19, 2026 03:09
8614108 to
80878a9
Compare
Contributor
There was a problem hiding this comment.
ARGUS
This PR changes 5,569 lines across 23 files. A change this size is hard to review
well, for a person and for this bot alike: reviewers lose the thread, defects hide in the
volume, and the feedback arrives too late to be cheap to act on. Splitting it into a
stacked pull request
gives each piece a reviewable diff and keeps the history. This review covers the whole change
regardless.
Bottom line: LGTM.
Comment @argus review to re-run, or @argus re-review after pushing.
dhanavikram
force-pushed
the
users/dhanavikram/net-helpers-host-tests
branch
from
October 19, 2026 15:01
80878a9 to
158bcc8
Compare
Contributor
There was a problem hiding this comment.
ARGUS
This PR changes 5,569 lines across 23 files. A change this size is hard to review
well, for a person and for this bot alike: reviewers lose the thread, defects hide in the
volume, and the feedback arrives too late to be cheap to act on. Splitting it into a
stacked pull request
gives each piece a reviewable diff and keeps the history. This review covers the whole change
regardless.
Bottom line: LGTM.
Comment @argus review to re-run, or @argus re-review after pushing.
dhanavikram
force-pushed
the
users/dhanavikram/net-helpers-host-tests
branch
from
October 19, 2026 17:51
158bcc8 to
a5da1f9
Compare
Contributor
There was a problem hiding this comment.
ARGUS
This PR changes 5,569 lines across 23 files. A change this size is hard to review
well, for a person and for this bot alike: reviewers lose the thread, defects hide in the
volume, and the feedback arrives too late to be cheap to act on. Splitting it into a
stacked pull request
gives each piece a reviewable diff and keeps the history. This review covers the whole change
regardless.
Bottom line: LGTM.
Comment @argus review to re-run, or @argus re-review after pushing.
dhanavikram
force-pushed
the
users/dhanavikram/net-helpers-host-tests
branch
from
October 20, 2026 05:09
a5da1f9 to
2cb1366
Compare
Co-authored-by: Dhanavikram Sekar <dhanavikram@users.noreply.github.com>
Drop this PR's copy of the hipMemGetHandleForAddressRange fake; develop now carries the same hook. Fake rcclNetRegSegmentCount, rcclNetGraphRegSegmentCount and rcclNetRegCommit in rccl_wrap_fakes, which net.cc's register paths call since #10885, and cover them. Build the Net target with _MICRO_DMABUF_EXPORT_DEF so standalone runs the cuMem DMA-BUF arms too.
dhanavikram
force-pushed
the
users/dhanavikram/net-helpers-host-tests
branch
from
October 26, 2026 19:27
c71fdc2 to
35d5f53
Compare
Contributor
There was a problem hiding this comment.
ARGUS
This PR changes 5,658 lines across 23 files. A change this size is hard to review
well, for a person and for this bot alike: reviewers lose the thread, defects hide in the
volume, and the feedback arrives too late to be cheap to act on. Splitting it into a
stacked pull request
gives each piece a reviewable diff and keeps the history. This review covers the whole change
regardless.
Bottom line: LGTM.
Comment @argus review to re-run, or @argus re-review after pushing.
1 task done
1 task done
pvallem
reviewed
Oct 26, 2026
Contributor
pr-rocm-systems-13548-oci-rccl-pr-smokeBuilds
FailuresNo failing tests in this run. 15 tests skipped. |
Contributor
pr-rocm-systems-13548-ruby64-rccl-pr-buildBuilds
FailuresNo failing tests in this run. 183 tests skipped. pr-rocm-systems-13548-ruby64-rccl-pr-smokeBuilds
Failures
15 tests skipped. |
Contributor
pr-rocm-systems-13548-oci-rccl-pr-smokeBuilds
FailuresNo failing tests in this run. 15 tests skipped. |
Contributor
pr-rocm-systems-13548-ruby64-rccl-pr-buildBuilds
FailuresNo failing tests in this run. 183 tests skipped. pr-rocm-systems-13548-ruby64-rccl-pr-smokeBuilds
Failures
15 tests skipped. |
Contributor
pr-rocm-systems-13548-ruby64-rccl-pr-smokeBuilds
FailuresNo failing tests in this run. 15 tests skipped. |
… of transport_net_ib_fakes Follow #13995 in net.cc: fake ncclIbSetListenGdrFlushFlag, ncclIbGetGpuFlushFlag, ncclIbGetGpuFlushFlagSeq and ncclIbSetGpuFlushFlagHint, give the fixture a p2p LL128 size, and cover the listen flag, the per-step sequence allocation and its gates, its publication instead of a flush, and its free. Move net.cc's other IB peers to their owners' fakes: netIbCast and rcclCastNetP2pPolicy to net_ib_cast_common/connect_fakes, and rcclNetP2pPolicy to transport_net_ib_p2p_policy_fakes, which targets that compile the real net_ib/connect.cc leave out. transport_net_ib_fakes.h forward-declares the IB types: init_fakes.h includes it ahead of init-test.cc's free() redirect.
dhanavikram
added a commit
that referenced
this pull request
Oct 27, 2026
…#14299) ## Motivation #13863 skips the Ruby and OCI coco smoke runs for PRs to develop that only touch `projects/rccl/test/host/*`. But a PR that adds or extends a micro host binary also has to register it outside `test/host/`, so it still books both clusters: - `projects/rccl/test/test_categories_micro*.yaml`: the binary's CTest categories, read by the in-tree `test/host` CMake. - `projects/rccl/tools/scripts/test_runner/configs/ci-precheckin.json` and `lib/test_executor.py`: math-ci precheckin and coverage registration. Open examples that run coco only because of these files: #13548, #13755, #13835, #13857, #13858, #13642, #13116. ## Technical Details `rccl_change_sets.py` `COVERAGE_RULES`, develop only: - **"host-only unit tests"** also covers `projects/rccl/test/test_categories_micro.yaml` and `test_categories_micro_*.yaml`. The GPU suites' `test_categories.yaml`, `_fixtures`, `_fixtures_debug`, `_mpi`, `_amdsmi` and `_altrsmi` do not match and still trigger coco, as would a future `test_categories_microbench.yaml`. - **New "test runner" rule**: `projects/rccl/tools/scripts/test_runner/*`. This skips coco for any change confined to `test_runner/`, not only registrations. `rccl-host-unit-tests.yml` runs the test_runner unit tests on every PR to develop touching `projects/rccl/`. `ci-precheckin.json` itself is exercised only by math-ci precheckin. It is a separate rule from host-only tests so each keeps its own examples and justification. Why coco loses nothing here: - Its PR jobs in ROCm/tRCCL (`ruby64-`/`oci-rccl-pr-build`, then `-pr-smoke`) run harness-owned `smoke-test.yaml` and `rccl-perf-regression.yaml`, plus `pr-device-api.yaml` on Ruby: GPU `rccl-UnitTests`/`Fixtures`/`Gdr` gtests and rccl-tests. - Their `install.sh -t` build does configure `test/host` in-tree and parse the micro category files. A parse error there is only a CMake warning, and TheRock Multi-Arch CI (a required check, no path filter) runs the same in-tree configure. - No PR job runs the micro binaries, CTest labels, or anything under `tools/scripts`. No committed tRCCL job names a test_runner JSON, and the harness never executes `test_executor.py`. Also corrects #13863's comment that the in-tree `test/host` build is compiled only by TheRock (coco's build compiles it too), and updates the `rccl-host-unit-tests.yml` header for the second rule. Unchanged: TheRock Multi-Arch CI, math-ci precheckin, and the label-gated AINIC TW / dispatch-only MI350P workflows that use the test runner. ## Issue Tracking JIRA ID: N/A. Follow-up to #13863. ## Test Plan - `python3 -m pytest .github/scripts/tests/*rccl*_test.py` (what `rccl-cluster-runner-allowlist.yml` runs). - New cases: registration paths skip on develop; GPU category files still match; a full new-micro-binary file list emits `rccl=false`; look-alikes (`test_categories.yaml`, `test_categories_micro.yml`, `test_categories_microbench.yaml`, `test_runner_x/`, `tools/test_runner/`) stay uncovered. - Mutation check on a scratch copy: reverting, dropping or widening each pattern, or pointing a rule at another workflow, fails the tests. - Ran `coco_paths` on the changed files of the 7 PRs above and on every `test/test_categories*.yaml` on develop. ## Test Result - 120 passed, 149 subtests passed. - All 7 PRs above now skip coco; 42 micro category files are covered and the 6 GPU category files still trigger coco. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/rocm-systems/blob/develop/CONTRIBUTING.md.
Move rcclNetP2pPolicy and the four GDR flush-flag accessors net.cc calls into transport_net_ib_connect_net_fakes, which targets that compile the real net_ib/connect.cc leave out. Cover the flag path's empty-step and no-flush gates and the reset after a failed sequence allocation, and name what each net_ib fakes file holds in the README.
# Conflicts: # projects/rccl/test/host/fakes/rccl_wrap_fakes.h
Contributor
There was a problem hiding this comment.
ARGUS
This PR changes 6,053 lines across 28 files. A change this size is hard to review
well, for a person and for this bot alike: reviewers lose the thread, defects hide in the
volume, and the feedback arrives too late to be cheap to act on. Splitting it into a
stacked pull request
gives each piece a reviewable diff and keeps the history. This review covers the whole change
regardless.
Bottom line: LGTM.
Comment @argus review to re-run, or @argus re-review after pushing.
dhanavikram
added a commit
that referenced
this pull request
Oct 27, 2026
- fakes/transport_net_ib_fakes.{h,cc}: develop (#13995) now carries this
file for the Init and Diagnostics binaries; keep its GDR flush-flag
seams and add this branch's connect.cc fakes. The header takes the same
forward-declaration form as #13548, so init_fakes.h keeps working.
- ncclNetIb moves to the new fakes/transport_net_ib_common_fakes.cc, owned
by common.cc. Init links it; Transport compiles the real common.cc and
leaves it out.
- #13995 points the ncclNetIb vtable at ncclIbNetListen (connect.cc), so
the connect.cc fakes gain a fail-loud ncclIbNetListen.
The RCCL host unit tests workflow already runs it; host-only microtests are leaving ci-precheckin.json.
Contributor
There was a problem hiding this comment.
ARGUS
This PR changes 6,036 lines across 27 files. A change this size is hard to review
well, for a person and for this bot alike: reviewers lose the thread, defects hide in the
volume, and the feedback arrives too late to be cheap to act on. Splitting it into a
stacked pull request
gives each piece a reviewable diff and keeps the history. This review covers the whole change
regardless.
Bottom line: LGTM.
Comment @argus review to re-run, or @argus re-review after pushing.
dhanavikram
added a commit
that referenced
this pull request
Oct 28, 2026
## Motivation `src/transport/coll_net.cc` (the CollNet transport: setup, connect, proxy progress, buffer registration and the comm-level CollNet setup) had no host-only tests, so its control paths were only exercised by runs on CollNet hardware. ## Technical Details Stacked on #13548 (`net.cc` host tests), which adds the topology and DMA-BUF seams this PR reuses; it will be rebased onto `develop` once that lands. New test TU `coll-net-test.cc` in the existing `rccl-UnitTestsMicro` binary. It `#include`s the hipified `coll_net.cc` and drives it through a fake `ncclCollNet_t`. Fakes: - `RCCL_STUBS_OMIT_COLL_NET` fences the CollNet stubs in `transport_stubs.cc` (`ncclCollNetSetup`) and `collective_stubs.cc` (the chain/direct buffer setups), since `rccl-UnitTestsMicro` now compiles the real ones. - `transport_stubs.cc` gains the GDRCopy params (pinned to their `net.cc` defaults) and seams for `ncclTransportP2pConnect`, the CollNet setup/check/free, `initCollnetChannel` and `ncclTransportInitRankMap` (default unchanged). - `rccl-UnitTestsMicro` now also links `rocmwrap_fakes.cc` and `topo_stubs.cc`. Covered: - `canConnect`, the DMA-BUF mapping flags, `collNetDumpMap`, the `collNetTransport` table - `send/recvProxySetup` and the shared collective comms and buffer pool (`sharedListen`, `sharedConnect`, `sharedFree`, `sharedBuffersInit/Destroy`) - `sendSetup`/`recvSetup`, `sendConnect`/`recvConnect`, `sendFree`/`recvFree` - `send/recvProxyConnect` (host or device pool, GDRCOPY, HSA and cuMem DMA-BUF with their fallbacks) and `send/recvProxyFree` - buffer registration: `ncclCollnetLocalRegisterBuffer`, `ncclCollnetGraphRegisterBuffer`, `ncclCollnetDeregBuffer`, `send/recvProxyRegBuffer`, `send/recvProxyDeregBuffer` - `calcAlgoOffset`, `calcRegionOffset` and the six collective issue helpers - `sendProxyProgress`, `collNetRecvFlush` and `recvProxyProgress` - `ncclCollNetSetup`, `ncclCollNetChainBufferSetup`, `ncclCollNetDirectBufferSetup`, `isCollNetChainHeadsSubset` The new suites are registered with one `CollNet*` pattern in `test_categories_micro.yaml`, and `MICROTEST_README.md` lists the unit. ## Issue Tracking JIRA ID: AICOMRCCL-2802 ## Test Plan - `run_host_tests.sh` (all phases). - `rccl-UnitTestsMicro --gtest_filter='CollNet*'` with `--gtest_shuffle --gtest_repeat=3`. - Also built in-tree (`BUILD_TESTS=ON`), which compiles the uncached-memory and cuMem DMA-BUF arms the standalone build leaves out. - Each commit builds and passes `rccl-UnitTestsMicro` on its own. - Mutation testing: changed the hipified `coll_net.cc` in about 1,140 places per build mode (standalone, in-tree and an ASan build) to check the tests catch each change. ## Test Result - All host binaries pass standalone and in-tree; `rccl-UnitTestsMicro` runs 2163 tests standalone (206 CollNet) and 2180 in-tree (214 CollNet). - 99.3% line coverage of `coll_net.cc`; the 10 lines not run are unreachable (e.g. the `fail:` fd closes, where the fd is always already -1). - Every surviving code change is equivalent (log text, a memory fence, a redundant bound), compiled out of that build, unreachable, or an out-of-bounds access only ASan sees. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/rocm-systems/blob/develop/CONTRIBUTING.md. --------- Co-authored-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Dhanavikram Sekar <dhanavikram@users.noreply.github.com>
ywang103-amd
pushed a commit
that referenced
this pull request
Oct 28, 2026
## Motivation `src/transport/net.cc` (the NET transport: setup, connect, proxy progress and shared buffer pools) had no host-only tests, so its control paths were only exercised by multi-node runs. ## Technical Details New `rccl-UnitTestsMicroNet` binary. Its test TU `#include`s the hipified `net.cc` and drives it through a fake `ncclNet_t`. It needs its own binary because `transport_stubs.cc`, which most targets link, defines `rcclUseAinic` and `netTransport`, and `net.cc` defines both. `net.cc`'s other externals become seams or fail-loud floors in the fakes file that owns each one (`topo_stubs`, `proxy_fakes`, `os_fakes`, `rocmwrap_fakes`, `transport_net_ib_fakes`, `transport_p2p_fakes`, `nccl_stubs`, `hip_fakes`). `hipMemGetHandleForAddressRange` is faked only under `RCCL_CUMEM_DMABUF_EXPORT_SUPPORTED`, which in-RCCL builds inherit from the `rccl` target. Covered: - IB NIC detection from a temporary sysfs tree, the network attribute hints, `canConnect`, flush-type names, `ncclNetGetDeviceHandle` - the shared p2p buffer pool (`sharedNetBuffersInit/Get/Destroy`, `proxySharedInit`) - `sendSetup`/`recvSetup`, `send/recvProxySetup`, `sendConnect`/`recvConnect`, `sendFree`/`recvFree` - `send/recvProxyConnect` (dedicated and shared buffers and comms, GDR, GDRCOPY, HSA and cuMem DMA-BUF, device offload) and `send/recvProxyFree` - buffer registration: `ncclNetLocalRegisterBuffer`, `ncclNetGraphRegisterBuffer`, `ncclNetDeregBuffer`, `send/recvProxyRegBuffer`, `send/recvProxyDeregBuffer` - `sendProxyProgress` and `recvProxyProgress` `rocmwrap_fakes.cc` now owns the `hsa_status_string` fake, so `rccl-UnitTestsMicroRmaProxy` (which already links it) drops its local copy. The binary is registered in `CMakeLists.txt`, `run_host_tests.sh`, `test_categories_micro_net.yaml` and `test_executor.py` (not `ci-precheckin.json`: host-only microtests run in the RCCL host unit tests workflow instead), and `MICROTEST_README.md` lists it. ## Issue Tracking JIRA ID: AICOMRCCL-2801 ## Test Plan - `run_host_tests.sh` (all phases) on develop f94e399. - `rccl-UnitTestsMicroNet` with `--gtest_shuffle --gtest_repeat=3`. - Also built in-tree (`BUILD_TESTS=ON`), which compiles the cuMem DMA-BUF arms the standalone build leaves out. - Mutation testing: changed the hipified `net.cc` in 524 places to check the tests catch each change. ## Test Result - All 20 host binaries pass standalone (Net and RmaProxy also checked in-tree before this rebase); Net: 270 tests standalone, 281 in-tree. - 516 of 524 code changes caught. The 8 left change nothing observable (e.g. dropping `setXferNetAttrs`'s null-plugin guard, which `setNetAttrs` repeats). - ~98% line coverage of `net.cc`. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/rocm-systems/blob/develop/CONTRIBUTING.md. --------- Co-authored-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Dhanavikram Sekar <dhanavikram@users.noreply.github.com>
dhanavikram
added a commit
that referenced
this pull request
Oct 28, 2026
transport_net_ib_fakes.cc: develop's #13548 copy of the flush-flag and GidInfoQuery fakes matches this branch, so only the fail-loud listen, connect, accept and close stubs are this branch's; ncclNetIb stays in transport_net_ib_common_fakes.cc, which rccl-UnitTestsMicroNet (net-test.cc reads and assigns ncclNetIb) now links. MICROTEST_README.md: one row per net_ib fake file, with #13548's connect_net_fakes row.
dhanavikram
added a commit
that referenced
this pull request
Oct 28, 2026
…and #14381 hsa_status_string: #13548 put a copy in rocmwrap_fakes.cc. As promised in review, hsa_fakes.cc stays the one owner; rccl-UnitTestsMicro (coll-net-test's rocmwrap.h HSACHECK) and rccl-UnitTestsMicroNet (net.cc), which used the rocmwrap copy, link hsa_fakes.cc instead. ncclParamDmaBufEnable: rocmwrap.cc owns it, so rocmwrap_fakes.cc keeps the fake and the copy develop added to dev_runtime_micro_fakes.cc goes; rccl-UnitTestsMicro links both files. rcclParamForceEnableDMABUF is develop's copy and pin. The DmaBufEnable pin is a _rccl_cast_pins row. connect-test.cc: ncclSocketClose comes from os_fakes after #13755.
Contributor
TheRock Submodule Bump ActivityNewest first
|
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
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.
Motivation
src/transport/net.cc(the NET transport: setup, connect, proxy progress and shared buffer pools) had no host-only tests, so its control paths were only exercised by multi-node runs.Technical Details
New
rccl-UnitTestsMicroNetbinary. Its test TU#includes the hipifiednet.ccand drives it through a fakencclNet_t. It needs its own binary becausetransport_stubs.cc, which most targets link, definesrcclUseAinicandnetTransport, andnet.ccdefines both.net.cc's other externals become seams or fail-loud floors in the fakes file that owns each one (topo_stubs,proxy_fakes,os_fakes,rocmwrap_fakes,transport_net_ib_fakes,transport_p2p_fakes,nccl_stubs,hip_fakes).hipMemGetHandleForAddressRangeis faked only underRCCL_CUMEM_DMABUF_EXPORT_SUPPORTED, which in-RCCL builds inherit from therccltarget.Covered:
canConnect, flush-type names,ncclNetGetDeviceHandlesharedNetBuffersInit/Get/Destroy,proxySharedInit)sendSetup/recvSetup,send/recvProxySetup,sendConnect/recvConnect,sendFree/recvFreesend/recvProxyConnect(dedicated and shared buffers and comms, GDR, GDRCOPY, HSA and cuMem DMA-BUF, device offload) andsend/recvProxyFreencclNetLocalRegisterBuffer,ncclNetGraphRegisterBuffer,ncclNetDeregBuffer,send/recvProxyRegBuffer,send/recvProxyDeregBuffersendProxyProgressandrecvProxyProgressrocmwrap_fakes.ccnow owns thehsa_status_stringfake, sorccl-UnitTestsMicroRmaProxy(which already links it) drops its local copy.The binary is registered in
CMakeLists.txt,run_host_tests.sh,test_categories_micro_net.yamlandtest_executor.py(notci-precheckin.json: host-only microtests run in the RCCL host unit tests workflow instead), andMICROTEST_README.mdlists it.Issue Tracking
JIRA ID: AICOMRCCL-2801
Test Plan
run_host_tests.sh(all phases) on develop f94e399.rccl-UnitTestsMicroNetwith--gtest_shuffle --gtest_repeat=3.BUILD_TESTS=ON), which compiles the cuMem DMA-BUF arms the standalone build leaves out.net.ccin 524 places to check the tests catch each change.Test Result
setXferNetAttrs's null-plugin guard, whichsetNetAttrsrepeats).net.cc.Submission Checklist