[Triton] Declare triton>=3.6.0 dependency

[micmelesse]

Motivation

This PR declares aiter's dependency on triton in setup.py. This is standard practice for python libraries. We declare a minimum version of 3.6.0. We remove a lot of ad-hoc code that was branching on triton versions.

Technical Details

Test Plan

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:triton-355	Run Triton tests on MI355 in addition to MI325
ci:sglang	SGLang integration tests
ci:atom	ATOM benchmark (DeepSeek-R1 + GPT-OSS)
ci:vllm	vLLM benchmark
ci:all	All of the above

Add labels via the sidebar or gh pr edit 2695 --add-label <label>

[Copilot]

Pull request overview

This PR formalizes Triton as an explicit runtime dependency (minimum 3.6.0) and simplifies the Triton/Gluon codepaths by removing legacy version-branching and compatibility shims across kernels and tests.

Changes:

• Add triton>=3.6.0 to install_requires in setup.py.
• Remove Triton version-detection helpers and conditional branches targeting older Triton versions across Gluon/Triton kernels.
• Update Triton op tests to drop Triton 3.5-specific logic/skips and disable AOT usage where incompatible.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
setup.py	Declares Triton (>=3.6.0) as a required dependency for non-Windows installs.
op_tests/triton_tests/gemm/basic/test_gemm_afp4wfp4.py	Removes Triton-3.5 AOT gating; forces use_aot=False for preshuffle path.
op_tests/triton_tests/gemm/basic/test_gemm_a8w8.py	Removes Triton-3.5-specific skip for FP8 split-K lowering.
aiter/ops/triton/gluon/triton_version.py	Deletes version parsing/constants module now made obsolete by the min-version requirement.
aiter/ops/triton/gluon/pa_decode_gluon.py	Removes version-dependent MFMA/logits layout branching; standardizes on >=3.6 layout/shape.
aiter/ops/triton/gluon/gemm_afp4wfp4.py	Removes version-dependent MFMA instruction shape selection.
aiter/ops/triton/gluon/gemm_a8w8.py	Removes version-dependent MFMA shape/width branching; assumes >=3.6 MFMA shapes.
aiter/ops/triton/attention/pa_mqa_logits.py	Drops legacy Triton version branching and unconditionally uses Gluon ASTSource path consistent with min-version.
aiter/ops/triton/_triton_kernels/gated_delta_rule/gated_delta_rule_utils.py	Removes outdated “recommended Triton>=3.2” warning logic.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[brunomazzottiamd]

@micmelesse, it LGTM! I've left a AOT cleanup suggestion from other PR, apply it at your own judgement.

[brunomazzottiamd]

Test failure report:

Triton Tests (MI35X) / Shard 0 - It's a know issue that's under investigation, it isn't a blocker for merging.

op_tests/triton_tests/attention/test_mha.py::test_mha_backward[True-False-0.0-True-128-8-32-512-512-1]
op_tests/triton_tests/attention/test_mha.py::test_mha_backward[True-False-0.0-True-128-8-32-512-512-4]
op_tests/triton_tests/attention/test_mha.py::test_mha_backward[True-False-0.0-True-128-8-32-1024-1024-1]
op_tests/triton_tests/attention/test_mha.py::test_mha_backward[True-False-0.0-True-128-8-32-1024-1024-4]
op_tests/triton_tests/attention/test_mha.py::test_mha_backward[True-False-0.0-True-128-8-32-2048-2048-1]
op_tests/triton_tests/attention/test_mha.py::test_mha_backward[True-False-0.0-True-128-8-32-2048-2048-4]
op_tests/triton_tests/attention/test_mha.py::test_mha_backward[True-False-0.0-True-128-8-64-512-512-1]
op_tests/triton_tests/attention/test_mha.py::test_mha_backward[True-False-0.0-True-128-8-64-512-512-4]
op_tests/triton_tests/attention/test_mha.py::test_mha_backward[True-False-0.0-True-128-8-64-1024-1024-1]
op_tests/triton_tests/attention/test_mha.py::test_mha_backward[True-False-0.0-True-128-8-64-1024-1024-4]
op_tests/triton_tests/attention/test_mha.py::test_mha_backward[True-False-0.0-True-128-8-64-2048-2048-1]
op_tests/triton_tests/attention/test_mha.py::test_mha_backward[True-False-0.0-True-128-8-64-2048-2048-4]

Triton Tests (MI325) / Shard 5 and Triton Tests (MI35X) / Shard 5 - It's a recently discovered issue caused by rocm/pytorch:latest update to 2.10.0+rocm7.2.2.git40d237bf. It's a test issue, related to torch.topk API, will be addressed soon and it isn't a blocker for merging.

op_tests/triton_tests/moe/test_moe_routing_sigmoid_top1_fused.py::test_routing_sigmoid_top1[dtype0-16-128-128]
op_tests/triton_tests/moe/test_moe_routing_sigmoid_top1_fused.py::test_routing_sigmoid_top1[dtype0-16-128-1024]
op_tests/triton_tests/moe/test_moe_routing_sigmoid_top1_fused.py::test_routing_sigmoid_top1[dtype0-16-128-2048]
op_tests/triton_tests/moe/test_moe_routing_sigmoid_top1_fused.py::test_routing_sigmoid_top1[dtype0-16-128-4096]
op_tests/triton_tests/moe/test_moe_routing_sigmoid_top1_fused.py::test_routing_sigmoid_top1[dtype0-16-128-8192]
op_tests/triton_tests/moe/test_moe_routing_sigmoid_top1_fused.py::test_routing_sigmoid_top1[dtype0-128-128-128]
op_tests/triton_tests/moe/test_moe_routing_sigmoid_top1_fused.py::test_routing_sigmoid_top1[dtype0-128-128-1024]
op_tests/triton_tests/moe/test_moe_routing_sigmoid_top1_fused.py::test_routing_sigmoid_top1[dtype0-128-128-2048]
op_tests/triton_tests/moe/test_moe_routing_sigmoid_top1_fused.py::test_routing_sigmoid_top1[dtype0-128-128-4096]
op_tests/triton_tests/moe/test_moe_routing_sigmoid_top1_fused.py::test_routing_sigmoid_top1[dtype0-128-128-8192]

Triton Tests (MI35X) / Shard 7 - Fixed by #2723, it isn't a blocker for merging.

op_tests/triton_tests/fusions/test_fused_mul_add.py::test_mul_add[dtype2-False-b_type_is_scalar1-a_type_is_scalar0-shape3]
op_tests/triton_tests/fusions/test_fused_mul_add.py::test_mul_add[dtype2-False-b_type_is_scalar1-a_type_is_scalar0-shape5]

@micmelesse, can you please check what's going on with Flash Attention Integration / Flash Attention - Triton / MI35X (1 GPU) job? If we figure this out then we're good to merge.

FYI: @azaidy

[brunomazzottiamd]

op_tests/triton_tests/moe/test_moe_routing_sigmoid_top1_fused.py is fixed by #2750.

[micmelesse]

@brunomazzottiamd I will rebase and fix the Flash Attention test failure.