Add gelu_tanh activation to no-quant CK 2-stage fused MoE

[jonahbernard]

Motivation

DiffusionGemma / Gemma4 MoE models use the tanh approximation of GELU (gelu_tanh) in their expert MLPs. The AITER fused 2-stage CK MoE kernel previously supported only silu and gelu (erf) activations, so on ROCm these models either fell back to the slower unfused Triton implementation or could not run the fused path with the correct activation. This PR adds gelu_tanh as an activation in the bf16 (no-quant) CK 2-stage MoE path so Gemma MoE can use the fused kernel with the mathematically correct activation.

Technical Details

Paired with this PR into CK: ROCm/rocm-libraries#8886

• Codegen (csrc/ck_gemm_moe_2stages_codegen/): added gelu_tanh to the activation tables. ACT_TO_INT = {"gelu": 0, "silu": 1, "gelu_tanh": 3} and matching INT_TO_ACT; widened
  ActOP from bool to int. In gen_instances.py, the no-quant loop now iterates acts + ["gelu_tanh"] so the new instance is generated only for the f16/b16 no-quant families (it stays out of the quant families). gelu_tanh added to the -act CLI choices.
• Enum remap (gemm_moe_ck2stages.cu): AITER's ActivationType enum (Silu=0, Gelu=1, Swiglu=2, GeluTanh=3) differs from CK's Activation enum (gelu_and_mul=0, silu_and_mul=1, …, gelu_tanh_and_mul=3). Replaced the old activation = !activation bool flip with an explicit aiter_act_to_ck() remap (Silu 0→1, Gelu 1→0, GeluTanh 3→3) at both stage1 and stage2 dispatch.
• Python op map (aiter/ops/moe_op.py): added ActivationType.GeluTanh: "gelu_tanh" to act2str_dict.
• Torch reference (aiter/ops/quant.py, get_torch_act): added GeluTanh → F.gelu(x, approximate="tanh") so the test oracle matches the CK FastGelu.
• Arg parsing (aiter/utility/dtypes.py, str2ActivationType): fixed snake_case→CamelCase parsing so -a gelu_tanh resolves to ActivationType.GeluTanh (previously "gelu_tanh".capitalize() → "Gelu_tanh" → AttributeError).

Test Plan

Verified on a gfx950 box using the standalone CK 2-stage MoE accuracy test (op_tests/test_moe_2stage.py), no-quant bf16, GeluTanh activation. The test compares the fused AITER/CK
kernel against a torch reference using F.gelu(x, approximate="tanh").

 python3 op_tests/test_moe_2stage.py \
  -q 0 -a gelu_tanh -d bf16 -e 32 -k 5 -t 64 -dim 512,256 --no-flydsl-csv

• -q 0 = no-quant (a16w16), the path this PR adds gelu_tanh to.
• -e 32 -k 5 keeps topk ≤ experts.

Test Result

shape (model_dim, inter_dim)	logits_diff	result
512, 256	6.99e-06	pass

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 on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 3972 --add-label <label>