[ariscript]

Stacked on #617 (review only the last commit). With these, plans can express reduce-scatter inboxes, all-gather into a staging buffer (as
AG-GEMM needs), and two-shot all-reduce.

What changes

• TriggeredPlan(destination=...) lands peers' chunks in another symmetric tensor.
  Region(dst_offset_bytes=, dst_stride_bytes=) places each region there at its own offset and row
  pitch. The schedule builders take dst= / dst_base=.
• Chunk(peers=...) sends a chunk to only some ranks. Other ranks neither receive it nor wait for it.
• Chunk(group=...): a group's chunks signal one arrival counter at each receiver.
  TriggeredView.wait_group() / group_arrived() wait for all of them at once.
• The overlap check runs per receiver, on destination bytes. When the destination is the buffer
  itself, what lands on a rank must also miss the bytes that rank sends; that is the previous rule,
  generalized.
• Each chunk record names the gate word that shows it on this rank and how much that word grows per
  round. arrived() and plan.wait() therefore treat own, grouped, ungrouped and not-addressed
  chunks alike.
• Not supported: sending a chunk to its own rank.

Testing (MI300X, torch allocator)

• test_plan.py at 2 and 4 ranks. New tests:
  • tiles all-gathered into a staging buffer with a wider row pitch;
  • a reduce-scatter-shaped exchange into per-source inbox slots, with one wait_group in the
    consumer kernel;
  • a chunk sent to one peer only.
• Schedule unit tests: destination offsets in the builders, validation, the per-receiver overlap check.
• View tests: recipients and groups.
• Docs build.

[Copilot]

Warning

Copilot couldn't run its full agentic review because it didn't start before the timeout. Make sure your repository has a runner available, or add a copilot-code-review.yml file specifying one with the runs-on attribute. See the docs for more details.

Copilot review overview

5 open issues

• dst_base being optional here can silently compute an incorrect destination offset when dst is a… · New
• Same issue as the Triton TriggeredView docstring: the text implies group semantics apply… · New
• The limit check now covers more than writers (it also incorporates grouped arrivals via… · New
• This docstring is inaccurate for the owner == rank case when chunk.group is set: view_words()… · New
• bind_sdma_ep() is called inside the inner loop (and chain_bytes() also binds internally), which… · New

What changed in this PR

This PR extends triggered plans to support non-default destination placement, selective peer delivery, and grouped arrivals to enable patterns like reduce-scatter inboxes, all-gather into staging buffers, and two-shot all-reduce.

Changes:

• Add destination placement (TriggeredPlan(destination=...), Region(dst_offset_bytes=..., dst_stride_bytes=...)) and propagate it through schedule builders and SDMA arming.
• Add selective delivery (Chunk(peers=...)) and grouped arrivals (Chunk(group=...), TriggeredView.group_arrived() / wait_group()).
• Update disjointness validation to check destination bytes per receiver (and protect own sent bytes when destination is the buffer).

File	Description
tests/​unittests/​test_plan_view_gluon.py	Updates view gate partitioning and adds tests for recipients/groups in Gluon.
tests/​unittests/​test_plan_view.py	Updates view gate partitioning and adds tests for recipients/groups in Triton view.
tests/​unittests/​test_plan_schedule.py	Adds unit tests for destination defaults, builders landing in dst, dst pitch, peer filtering, and per-receiver disjointness checks.
tests/​unittests/​test_plan.py	Adds integration tests for tiling into staging w/ pitch, reduce-scatter inbox + wait_group, and chunk-to-some-peers.
iris/​experimental/​plan/​triton.py	Bumps ABI, expands view record layout, and implements group arrival/wait logic.
iris/​experimental/​plan/​triggered.py	Adds destination support, per-peer queue sizing, group gate allocation, updated view encoding, and per-receiver overlap validation.
iris/​experimental/​plan/​schedule.py	Extends Region with destination placement, Chunk with peers/groups, updates builders, validate(), and check_disjoint() logic.
iris/​experimental/​plan/​kernels.py	Updates wait_kernel to use per-chunk arrival word + expected-per-round derived from records.
iris/​experimental/​plan/​gluon.py	Mirrors Triton TriggeredView changes (arrival word/expected + group wait APIs).
iris/​experimental/​plan/​armer.py	Adds dst_pitch to Copy, updates per-peer queue sizing for selective peers, and uses destination pitch for tiled copies.
docs/​reference/​host/​plan.md	Exposes group_arrived/wait_group in API reference.
docs/​conceptual/​triggered-plans.md	Documents destinations/peers/groups and updates constraints/limits accordingly.

_{Review effort: Lite}