[TRITON] Add act_mul without quant (DO_QUANT), model configs, benchmarks

[Chi-Chu319]

• Fix act_mul in aiter/ops/triton/activation.py to handle higher-rank inputs ([..., 2*d]) by flattening leading dims before the kernel call and restoring the original batch shape on return
• Update docstring to document the reshape behavior
• Fix torch_act_mul_ref in tests to use ... indexing for higher-rank support
• Add test_act_mul_no_quant_higher_rank test covering 3D and 4D inputs
• Run parallel validation

[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 2592 --add-label <label>

[Copilot]

Pull request overview

This PR extends the Triton activation/gating path by adding a non-quantized act_mul API that reuses the existing quantization-capable kernels with quant disabled, plus associated tests, benchmarking support, and new model benchmark configs.

Changes:

• Add act_mul (no quantization) in aiter.ops.triton.activation, implemented by calling the existing FP8-group-quant kernel with DO_QUANT=False.
• Add unit tests for the new no-quant act_mul behavior (including out= support) and add a dedicated -bench_act_mul mode to the MoE benchmark script.
• Add GLM-4.7 and Kimi-K2.5 model entries to the Triton benchmark model config JSON.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
aiter/ops/triton/activation.py	Adds act_mul and wires DO_QUANT into kernel launches.
aiter/ops/triton/_triton_kernels/activation.py	Adds DO_QUANT constexpr paths to support quantized vs non-quantized stores.
op_tests/triton_tests/test_activation.py	Adds correctness tests for act_mul (no quant).
op_tests/op_benchmarks/triton/bench_moe.py	Adds -bench_act_mul benchmarking mode and activation selection.
op_tests/op_benchmarks/triton/utils/model_configs.json	Adds glm47fp8 and kimik25 benchmark configurations.

---

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

[juuso-oskari]

LGTM

[ChuanLi1101]

LGTM.

The new act_mul no-quant entrypoint and the DO_QUANT constexpr addition look clean. Tests cover 2D and higher-rank inputs (3D/4D), three activations (silu / gelu / gelu_tanh), and the with/without out= paths. CI looks healthy aside from one Triton MI35X Shard 7 flake -- re-ran the failing shard, looks like the same GPU-oversubscription pattern Ali (@azaidy) called out on #2902.

@vgokhale your feedback is reasonable -- esp. the redundancy of DO_QUANT on the fp4 kernel given the new wrapper only drives the fp8 path. @Chi-Chu319 could you address the three points in a follow-up commit (drop DO_QUANT from _act_mul_and_dynamic_mxfp4_quant_kernel, trim the float16/float32 cases from the higher-rank test) so we keep the merge window open? Approving on the basis that the concern is a polish item rather than a correctness issue.

[brunomazzottiamd]

Please check if my comments in #3168 (review) are applicable to this PR. Thanks!