Repository navigation
Conversation
… dev/dasidler/sdma
… dev/dasidler/sdma
… dev/dasidler/sdma
… dev/dasidler/sdma
…to dev/dasidler/sdma-benchmark
…rologueConfig - tritonblas: pin to the tritonBLAS branch that merges main (SignalView/WaitView, EpilogueConfig/PrologueConfig) into muhaawad/iris, installs origami from PyPI (rocm-origami), and includes the gfx950 LDS estimate fix. - rocm-xio: pin to current main (97fa166), 24 commits past the previous pin. - iris.ops copy-engine variants: construct EpilogueConfig / PrologueConfig directly and pass epilogue_config / prologue_config to persistent_matmul_lt, replacing the removed create_counter_config / create_wait_config helpers. Co-Authored-By: Claude <noreply@anthropic.com>
dsidler
changed the title
dsidler
commented
Oct 23, 2026
| "requests", | ||
| "ruff<0.17", # 0.16 formats code blocks in Markdown; see lint.yml | ||
| "tritonblas @ git+https://github.com/ROCm/tritonBLAS.git@muhaawad/iris", | ||
| "tritonblas @ git+https://github.com/ROCm/tritonBLAS.git@dev/dasidler/merge-main-into-muhaawad-iris", |
Collaborator
Author
There was a problem hiding this comment.
revert once the two tritonBLAS PR merged:
ROCm/tritonBLAS#100
ROCm/tritonBLAS#101
dsidler
marked this pull request as ready for review
October 23, 2026 22:00
dsidler
requested review from
BKP,
artulab,
drprajap,
efric,
mawad-amd,
neoblizz and
nirvedhmeshram
as code owners
October 23, 2026 22:00
This branch has not been deployed
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Adds copy-engine (SDMA) variants of the fused all-gather + GEMM ops. The SDMA engines move the remote tiles while the compute units run a persistent tritonBLAS GEMM, so no workgroups are spent on fetching.
What's in it
Ops (
iris.ops)all_gather_matmul_copy_engine: host-initiated or device-initiated (device_initiated=True) transfers. Remote K-shards are posted in M-batches, and the GEMM waits per batch onPrologueConfigflags.matmul_all_gather_copy_engine(device-initiated) andmatmul_all_gather_host_copy_engine: the GEMM signals tile completion throughEpilogueConfigcounters, and SDMA gathers launch-wave by launch-wave (tritonblas_launch_wave_schedule.py).iris.iris(..., engines=("sdma",))and fail fast with a clear error otherwise.Host and device API
Iris.put_tile/put_tilestake strided 2D copies, optional signal flags and wait-then-put.iris.mem.triton:put_signal,put_signal_rect,wait_then_put_rect(s),wait_then_put_signal_rect(s).Benchmarks and tests
torch.mm/ tritonBLAS + RCCL for matmul + all-gather.iris.benchhas a new--engines sdmaflag. The copy-engine benchmarks skip with a message without it.IRIS_TEST_*env vars. The default heap is 8 GB.Fixes to existing ops
Candstaged_ainall_gather_matmulandall_gather_matmul_hbm_buffer.all_gather_matmul_hbm_bufferhalves the defaultk_per_flaguntil it divides the K-blocks, so small and single-rank problems work.🤖 Generated with Claude Code