chore(runtime): pin the runtime submodule to simpler 4e4d3a4a

[Little-oil]

Summary

Move the runtime gitlink from 77fa0171 (main today, #2641) to 4e4d3a4a — simpler "CI: bump pinned pto-isa to a8040450" (#2123), 27 commits forward. This re-lands the range #2641 stepped back over (77fa0171..15f5cbd9) and adds what simpler has merged since.

The pin chain moves with it:

• pto-isa 96ba706c → a8040450. runtime/pto_isa.pin is the source of truth and toolchain/versions.env derives pto-isa from that file, so there is no pypto-side pin to edit. The one onboard-relevant fix in the range is 5d0fb9a8 "fix TCI A2A3 for prefill EP=2 NAN problem"; 07953421 corrects the tsort32 golden to the stable-quicksort order VBS32 actually produces, and 9c864fb4 reverts the TMATMUL_MX ping-pong change from earlier in the same range. The rest is CPU_SIM work (TLOAD padding, TMATMUL_ACC/TGEMV_ACC seeding, mixed-type TADD/TABS, TCI/TCVT scratch overloads, GM/UB transfer layout, DIR_BOTH FIFO order, FP4 ConvTile) plus A5 fixes (int64 shift carry, UB ND→NZ, TSel bf16).
• ptoas stays at v0.57.

Simpler commits entering the pin (77fa0171..4e4d3a4a, newest first):

simpler PR	Change
#2123	CI: bump pinned pto-isa to a8040450
#2070	elide A5 terminal releases after orchestration
#2099	dispatch-correlated same-host multi-rank swimlane timeline
#2117	docs: the device log threshold is fixed at device init
#2105	skip the empty Tier-0 staging order in the a5 dispatch loop
#2118	drain the host-log writer before scene tests read captured spans
#2112	rename maybe_rendezvous_ring to try_launch_sync_start_cohort
#2107	fix the swimlane pool layout at compile time
#2113	drop SimplerHostSpan's version and size words
#2110	classify Graph tensor sources by address, name them by kind
#2111	CI: bump pinned pto-isa to c0e28cfb
#2108	docs: why the host-log claim budget is not the knob it looks like
#2095	host_build_graph early dispatch via an ED publish list
#2029	host logging becomes nonblocking with accountable loss
#2101	host_build_graph includes say what each file uses
#2090	A5 HBG single-lane scheduling switches to AICore
#2100	AICPU scheduler timeout collapses to one 20 s constant
#2098	host_build_graph's host side stops reaching into device facilities
#2093	DFX collectors bind their per-run output separately
#2094	host_build_graph orchestration confined to the host
#2092	simpler_init provisions the async-DMA workspace itself
#2087	each runtime gets its own TaskId type
#2089	async-DMA provisioning asks for one flag, not an engine set
#2091	ProfilerBase::quiesce() drains without retiring the threads
#2088	docs for the async-DMA workspace
#2072	complete A5 HBG single-lane scheduling
#2086	one signed counting domain for host_build_graph

One source change: the per-dispatch DFX directory

simpler #2099 appends rank{r}/d{k} to output_prefix inside the ChipWorker child, for every diagnostic that writes below it. _submit_chip was appending the same suffix on the driver side, so under this pin the two nest and every artifact lands at <dfx_base>/rank{r}/d{k}/rank{r}/d{k}/ — two levels below where the post-pass looks. dist-system-tests caught it: test_l3_dfx_per_rank found the outer directory (this driver still stamps dispatch_program.json into it) and none of the files in it.

The driver now only derives that directory, for the dispatch-program marker, _clear_dfx_dispatch_dirs and the _collect_l3_swimlane globs, and forwards config untouched. TestSubmitChip asserted the mutation, so its cases now assert what is left: the base prefix reaches the submit un-suffixed, and the derived directory is named by where dispatch_program.json lands. The two counters stay in step because both count exactly the dispatches this driver marks: _make_call_config sets output_prefix only when a DFX flag is on, and the runtime splits only a config that carries one.

Known gap, recorded rather than papered over: a prepared DistributedWorker running the swimlane two-pass reuses its Worker, so the runtime's per-child capture counter does not reset alongside this driver's and pass 2 would land in d{k+1}. The one-shot path — which is what the ST covers — creates a fresh Worker per pass, resetting both. Pairing those captures by the dispatch_identity.json sidecar #2099 now writes is the durable fix and belongs in its own change.

Nothing else needed a source change

The other refactors that touch what pypto emits or includes were checked directly:

• [Bug] Column-slice of a dynamic-pitch tensor passed orch→incore silently reads wrong data (stride densified) #2087 split task_interface/task_id.h per runtime. Codegen emits TaskId and TaskId::invalid() into orchestration TUs (src/codegen/array_op_codegen.cpp); both are present in the new host_build_graph/task_id.h, which the orchestration include path already reaches. Nothing in pypto includes the moved headers by name.
• feat(distributed): host-orchestrated ring allreduce with builtin.tensor.allreduce_ring #2094 / feat(printer): add explicit-layout dump mode and PassDumpLevel enum #2101 moved RuntimeOps out of orchestration_api.h into runtime_ops.h and dropped its transitive <vector> / <array> / <mutex> / <thread> includes. Generated orchestration includes only orchestration_api.h and tensor.h (plus the standard headers codegen emits itself), and a probe TU built with the exact -I list from test_generated_orchestration_compiles_against_the_pinned_runtime syntax-checks clean — no repeat of the common/hierarchical/types.h shadowing that chore(runtime): bump the runtime submodule to simpler 15f5cbd9, renaming TaskTensor to Tensor #2617 had to work around.
• fix(ci): restore clang-tidy checks on main #2110 added is_graph_record_address and feat(ir): Add reinterpret_view dtype alias operator #2113/feat(ir): support odd automatic vector splits #2107/fix(ir): Validate row reduction scratch shape #2099 reworked the span and swimlane internals without touching orchestration_api.h, host_build_graph/types.h, tensor.h or task_id.h, so nothing pypto compiles against moved.
• GRAPH_MAX_TENSOR_ARGS is still 128, matching the kMaxBoundaryTensors constant in legalize_graph_boundary_pass.cpp and verify_graph_functions.cpp.
• CallConfig grows a capture_clock_anchors field (fix(ir): Validate row reduction scratch shape #2099). pypto constructs CallConfig() and sets named attributes through simpler's nanobind binding rather than mirroring the struct, so the wire-layout change is contained in the runtime.
• The only pypto include of a runtime header, platform_comm/comm_context.h from the collective kernel templates, is unaffected across this range.

One consumer-side note that is not a pypto change: #2029 adds _flush_host_log, _start_host_log_writer and the two host-log counters to simpler's _task_interface extension, so a local checkout needs its bindings rebuilt after this bump. pypto calls none of those symbols.

Test plan

• pypto_core builds clean against this pin, rebased onto refactor(runtime)!: store platform as its two axes, arch and execution_mode #2642 (cmake --build build --parallel 64, exit 0).
• Orchestration header probe compiles under the UT's include order (g++ -std=c++17 -fsyntax-only, exit 0).
• First CI run on this pin: 19/20 green; dist-system-tests failed on the nested DFX directory, fixed above.
• Second run: dist-system-tests green; TestSubmitChip still asserted the old contract, updated in the same commit.
• tests/ut/runtime + tests/ut/codegen locally: 1719 passed, 63 skipped.
• Fourth run: all 20 checks green.
• CI on the final (comment-only) revision.

One flake to flag rather than bury: test_l3_host_tensor_allreduce.py::test_host_tensor_allreduce_loop[2] failed once on the third run — round 2 read round 0's values (max diff = 20155.0, the offsets that test spaces its rounds by), i.e. the reused allreduce signal's self-clearing epilogue lost a race. The same tree passed it on the next run, and the run before it, so it is not deterministic. Worth watching under this pin all the same: simpler #2095 changes host_build_graph's dispatch timing, which is the kind of change that wakes a sleeping signal-reuse race. Three runs on this pin, one failure.

🤖 Generated with Claude Code

https://claude.ai/code/session_014EGTChvJLW5gXVDDMYzFdH

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Team

Run ID: c6ca4f2b-1bb2-4f54-a5a8-7901ca36a399

📥 Commits

Reviewing files that changed from the base of the PR and between b816516 and 54efe34.

📒 Files selected for processing (1)

• runtime

Included review availability: Your plan provides up to 2 included reviews per hour; 1 remains after this review.

---

📝 Walkthrough

Walkthrough

The runtime submodule pointer changes from one pinned commit to another. No exported or public entity declarations change.

Changes

Runtime revision update

Layer / File(s)	Summary
Update runtime submodule pointer runtime	The submodule reference changes to commit 4ec5f37dc5aa4e89d2032e8b2cf2848debe3e8da.

Estimated code review effort: 1 (Trivial) | ~2 minutes

Merge Risk: ⚪ Minimal · up to 54efe

This updates the pinned runtime revision while preserving the existing pypto source and ptoas version. The supplied build checks pass and no actionable merge-blocking risk is identified.

Poem

A rabbit checks the runtime pin
One fresh commit hops in
The pointer rests in place
No public names change their face
Carrots celebrate the revision grin

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 0…
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly identifies the main change: pinning the runtime submodule to a simpler revision. It is concise and technically relevant to the changeset.
Description check	✅ Passed	The description is directly related to the runtime submodule update. It explains the pin changes, affected runtime components, compatibility checks, tests, and known risks.

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}