## meta-pytorch/monarch — v0.5.0…v0.6.0

_426+ commits._

### Features
- **add Handle, an observe-only handle to a background Tokio task (#4363)** (11b8037)
- **Add telemetry self-message benchmark (#4361)** (802314d)
- **add --plaintext flag to talk plain HTTP (#4383)** (19e2fff)
- **add explicit channel completion sinks (#4261)** (b1ef226)
- **Add ActorRef Into impl for checked behavior view (#4366)** (3297d90)
- **Add ActorStatus::Zombie (#4351)** (c4db77d)
- **Add language tags to docs code blocks for syntax highlighting (#4358)** (c0ae7c9)
- **Introduce Child Actor Teardown Intent (#4349)** (affe7fc)
- **tp: Add `const_generics` to `small_vec`** (4c7d9d8)
- **ibverbs: introduce IbvMr and Mlx5dvMkey RAII wrappers (#4353)** (8768444)
- **ibverbs: introduce IbvPd, hold the PD as Arc<IbvPd> (#4352)** (a4594f5)
- **Add out-of-cluster support for kubernetes (#3153)** (71d66e1)
- **Add non zero usize newtype (#4346)** (832f19b)
- **add an out-of-band mesh-reference table to PythonMessage (#4330)** (94f4334)
- **python/tests: add deferred-pickle characterization tests (#4328)** (53bbbe1)
- **Add docstring for `endpoint` and document `concurrent_endpoint` (#4332)** (b5ea260)
- **Add snapshot ingest API to telemetry sidecar (#4158)** (299d23e)
- **Add snapshot batch ingest to `DatabaseScanner` (#4157)** (e3b1ddc)
- **Add MlxQueuePair and finish the mlx5 integration (#4319)** (e2b8e69)
- **Add the standard-ibverbs RCQueuePair (#4318)** (951552e)
- **Add IbvCq/IbvQp RAII handles and raw-qp helpers (#4317)** (7797b49)
- **Introduce the IbvQueuePair trait (#4316)** (e11a281)
- **Add `ProcSpawner` and `UnixProcSpawner` (#4312)** (45add41)
- **add Future._take_inner() and retire the _status.coro reach-ins (#4287)** (2950395)
- **python/tests: add endpoint-panic, sync, and supervisor-delivery characterization tests (#4286)** (26e1854)
- **python/tests: add explicit-response-port failure characterization test (#4284)** (2a642b3)
- **python/tests: add endpoint-exception supervision characterization tests (#4283)** (0ea479b)
- **python/tests: add Future state-machine characterization tests (#4282)** (1dc5856)
- **Support `Local` transport for duplex channels (#4291)** (7ea24b7)
- **SlurmJob: support SLURM batch mode (via `client_script`) (#4281)** (3f3c8cc)
- **add typed control port accessors (#4276)** (3f2a97f)

### Fixes
- **Fix macOS telemetry thread ID build (#4415)** (68d68e6)
- **Fix CQS signal modernize-use-using in fbcode/monarch/python (#4409)** (cdce05a)
- **fix: ProcessJob reaps detached worker session on kill Summary @allow-mixed-users-changes (#4398)** (4a1d68c)
- **Fix telemetry benchmark merge conflict (#4400)** (ba7c04a)
- **Fix Python actor queue dispatch under lazy imports (#4376)** (22f3f8e)
- **fix stale PyShared.__reduce__ docstring (#4364)** (fb517cf)
- **Fix for flaky sccache access to S3, fallback to rustc (#4326)** (e77d658)
- **fix the unsanctioned control-plane GIL test under mode/opt (#4338)** (7b7dc77)
- **Fix test_local_supervision_propagation topology race (#4336)** (3922a7f)
- **Fix test_config.py for bumped MESH_ATTACH_CONFIG_TIMEOUT default (#4311)** (700dded)
- **Fix debugger deadlock under queue dispatch (#4309)** (6f503c7)
- **Fix dial-router non-stale-close test hang (#4301)** (e5d2f9a)
- **Fix unused_imports lint in fbcode/monarch/hyperactor_mesh (#4296)** (5c58dd0)
- **Fix unittest XML uploading on failures (#4295)** (256c915)

### Backend
- **Bump socket2 0.6.4 -> 0.6.5** (d53ea9a)
- **Freeze uv usage in OSS build paths (#4408)** (8a4d54a)
- **cap mlx5dv segment MRs at device max_sge (#4410)** (41faf68)
- **scatter host-memory RDMA registrations across NICs (#4357)** (e51b4a3)
- **Bump uuid 1.23.4 -> 1.23.5** (8789192)
- **Make ActorStatus::Zombie part of ActorStatus::Stopping (#4369)** (d14ec51)
- **Use locked npm installs for dashboard builds (#4407)** (e2a5ea9)
- **migrate mesh-admin example demos to job.kill() teardown (#4399)** (f60f144)
- **Sync OSS lockfiles with fbsource vendored dependencies (#4401)** (5796371)
- **drive flush_async through Future so it actually flushes on an asyncio loop (#4387)** (a3a3c8b)
- **Bump stable PyTorch to 2.13 and nightly Monarch to 0.7 (#4397)** (fafb81c)
- **warn in the header when the terminal hides detail panes (#4393)** (0a68bd4)
- **hide the inbound-ordering pane when nothing is stalled (#4392)** (6c14193)
- **keep the flight recorder pinned to the newest events (#4391)** (75bc7b8)
- **Bump cc 1.2.66 -> 1.2.67** (c0be4f8)
- **Bump ignore 0.4.27 -> 0.4.28** (bb83ae4)
- **Bump bytes 1.12.0 -> 1.12.1** (d8359dc)
- **Bump regex 1.12.4 -> 1.13.0** (b379f23)
- **refuse Future.get() on a tokio thread, validate timeout before spawning (#4368)** (298f501)
- **make the public Future's asyncio surface Handle-backed (#4367)** (9345179)
- **Remove entity telemetry replay buffer (#4382)** (1de10df)
- **Rename job telemetry component (#4380)** (4f5b87f)
- **Remove direct snapshot telemetry ingest path (#4381)** (610e5d0)
- **Remove legacy distributed telemetry actor path in Rust (#4379)** (27391a4)
- **Remove legacy telemetry in python API (#4378)** (18b585a)
- **Bound ActorMesh supervision callbacks by rank count (#4386)** (a90197f)
- **mesh-admin: prove the API surfaces a queue-dispatch deadlock (#4385)** (9f00eb2)
- **make TLS setup fail-closed; honor the addr scheme (#4384)** (f6732a3)
- **delete PortReceiver's now-unused supervision-monitor plumbing (#4374)** (5d9d6c6)
- **drop the dead supervision-monitor branch in PortReceiver._recv (#4373)** (acfae41)
- **migrate airport turnaround demo to concurrent endpoints (#4372)** (81fe7b7)
- **pin airport turnaround demo to direct dispatch** (cfdac2d)
- **replace channel completion oneshots with receipts (#4264)** (52d54b3)
- **relax mailbox flush counters (#4263)** (51995b1)
- **optimize completion sink accepted path (#4262)** (30c40f0)
- **remotemount: RDMA data-plane optimization (leader-mediated RDMA fan-out) (#4356)** (de6ed01)
- **Bump cc 1.2.65 -> 1.2.66** (8ec16d8)
- **Cast proc_states via GetHostProcStates through the host tree (#4238)** (7854e95)
- **Surface attach/simplex distinction at the link layer (#3660)** (66a1606)
- **Move actor status writer to InstanceCell from InstanceState (#4350)** (a42b62d)
- **Propagate drain mode to child actors (#4306)** (fd8b802)
- **Bump ignore 0.4.26 -> 0.4.27** (b8a69e1)
- **Bump indicatif 0.18.4 -> 0.18.6** (f043b1c)
- **Bump humantime 2.3.0 -> 2.4.0** (79120b6)
- **Bump rand 0.10.1 -> 0.10.2** (df81042)
- **Bump arc-swap 1.9.1 -> 1.9.2** (da6c74b)
- **Bump memmap2 0.9.10 -> 0.9.11** (4e341aa)
- **Bump quinn 0.11.9 -> 0.11.11** (26a109f)
- **Bump uuid 1.23.3 -> 1.23.4** (a290fba)
- **Bump quote 1.0.45 -> 1.0.46** (fc60697)
- **Bump anyhow 1.0.102 -> 1.0.103** (7d7bcc1)
- **Bump which 8.0.3 -> 8.0.4** (722714b)
- **Bump bytes 1.11.1 -> 1.12.0** (52331fb)
- **Bump cc 1.2.64 -> 1.2.65** (2d19a95)
- **Move proc_states from HostMeshRef to ProcMeshRef (#4235)** (a819428)
- **Cast StreamState through the host-mesh tree (#4237)** (09425d3)
- **ibverbs: hold domains by value, not Arc<IbvDomain> (#4355)** (d96b3c7)
- **ibverbs: make IbvQp and IbvCq own their backing resources (#4354)** (ab5fb1f)
- **HostMesh cleanup (#4102)** (adcd6ed)
- **Configure Tokio worker thread count (#4347)** (64cd3ed)
- **carry out-of-band mesh references as typed multipart parts (#4345)** (0a83ee9)
- **python/tests: pin bare-pickle of mesh refs (resolved survives, unresolved blocks) (#4343)** (15af24f)
- **delete the now-unused deferred-pickle subsystem (#4337)** (3be85f4)
- **carry mesh references out-of-band, with ref-aware decode (#4335)** (dd436a3)
- **route mesh references through the refs table, gated off (#4331)** (8f27bd9)
- **Cast proc spawn with SpawnProcs (#4104)** (d3cf907)
- **treat max_rd_atomic == 0 as "no separate read limit" (#4342)** (e063c8f)
- **derive the RoCE v2 source GID from device info, not a config knob (#4334)** (ac9e855)
- **static test_rdma_action.py device variants (#4323)** (98a2b01)
- **Document key `monarch.actor` public API symbols (#4333)** (bb41d49)
- **Attach serve_via before minting host refs (#4322)** (1617670)
- **remotemount: on-demand per-block FUSE mount (simple per-worker transport) (#4324)** (6152a60)
- **Make default_bootstrap_cmd public API (#4327)** (d9fc488)
- **Ensure entire RPC path covered by debug spans (#4329)** (8e05f36)
- **Skip redundant local telemetry collectors (#4197)** (7a224f6)
- **Revert cargo dependencies override for reqwest (#4325)** (e942aa8)
- **Use `OnceCell` for runtime identity tagging (#4320)** (7b281e4)
- **Serialize live dashboard `QueryEngine` queries (#4174)** (ac3a095)
- **Make telemetry sidecar work on remote hosts (#4321)** (9cb1614)
- **Job sidecar to inherit the parent transport (#4170)** (98aee8d)
- **Flip examples to the opt-in telemetry sidecar (#4169)** (b808c26)
- **Remove distributed_telemetry.py and distributed_telemetry_mast.py (#4166)** (8351e96)
- **Publish snapshots over telemetry sidecar HTTP (#4163)** (700dc19)
- **Migrate telemetry tests to sidecar query client (#4155)** (9bf9485)
- **Wire telemetry sidecar into jobs (#4154)** (dcb1863)
- **close the control-plane GIL net: route remaining Python::attach + ban raw attach (#4314)** (c297ef0)
- **extract the GIL wrappers into a monarch_gil base crate (#4308)** (9953961)
- **flag unsanctioned GIL acquisition on the control-plane runtime (#4305)** (f32131a)
- **Move IbvQueuePair into a legacy submodule (#4315)** (a719356)
- **tag the control-plane and stream runtimes (RuntimeKind) (#4303)** (0a5dd2a)
- **Refactor job lifecycle into components (#4208)** (9d8eea0)
- **run the RDMA actors on their own data-plane runtime (#4302)** (b9687b7)
- **Flip `actor_queue_dispatch` default to True (#4211)** (e3db7c7)
- **re-apply transport config on unpickle for batch mode (#4297)** (b32f794)
- **python/tests: pin sync-endpoint loop-invisibility (fake_sync_state) (#4289)** (c94c915)
- **concurrent_endpoint: fail the actor on an escaped exception (#4300)** (05c3b56)
- **Bump MESH_ATTACH_CONFIG_TIMEOUT to 60s (#4299)** (e21fb3a)
- **python/tests: pin all-sync-or-all-async endpoint enforcement (#4298)** (77c2b94)
- **once_cell -> std::sync in fbcode/monarch** (6742447)
- **Update autocargo component on FBS:master** (7642aff)
- **Update hyperactor book documentation (#4293)** (f60e192)
- **Attach a client gateway to an in-mesh host via Gateway::serve_via (#4142)** (46347a1)
- **Register the whole CUDA allocation for dmabuf MRs (#4285)** (2c96269)
- **Cast HostMesh drop cleanup shutdown (#4106)** (68ffddb)
- **Replace lazy_static with std::sync::LazyLock in fbcode/monarch** (2abf737)

_Recap by [Repo Wrapped](https://repowrapped.com/gh/meta-pytorch/monarch?utm_source=github-action)._