fix(kimi): align input norm quant with attention quant

[qichu-yun]

Motivation

Fix input RMSNorm quant fusion gating for Kimi/DeepSeek-style attention paths. The previous logic could infer quantization from the decoder layer prefix instead of the actual attention input projection, enabling fused RMSNorm quantization even when that projection could not consume the quantized activation. The PR also keeps non-Triton FP8 support for Kimi while avoiding unsupported block/group FP8 paths such as GLM-5.1-FP8.

Technical Details

• Resolve the real attention input projection (fused_qkv_a_proj for LoRA-Q MLA, otherwise q_proj) and use its quant config to decide whether input RMSNorm quant fusion is allowed.
• Support online quant configs when the offline projection config does not accept quantized input.
• Enable fusion only for real FP8/FP4 quantized attention inputs.
• Allow non-Triton FP8 fusion only for per-token FP8, preventing unsupported FP8 block/group quant kernels from being used.
• Preserve the existing non-Triton MXFP4 gating path.

Test Plan

1. fix the bug:
launch server:

export HIP_VISIBLE_DEVICES=0,1,2,3
export AITER_QUICK_REDUCE_QUANTIZATION=INT4
export AITER_USE_FLYDSL_MOE_SORTING=1
export AITER_AR_1STAGE_MAX_KB=512

model_path=/shared/data/amd_int/models/Kimi-K2.5-MXFP4
vllm serve $model_path \
    --host localhost \
    --port 9000 \
    --tensor-parallel-size 4 \
    --trust-remote-code \
    --async-scheduling \
    --gpu_memory_utilization 0.9 \
    --kv-cache-dtype fp8 \
    --max-num-batched-tokens 16384 \
    --max-model-len 16384 \
    --compilation-config '{"cudagraph_mode": "FULL_AND_PIECEWISE"}' \
    --profiler-config '{"profiler": "torch", "torch_profiler_dir": "kimi_k2_5_mxfp4_attnfp8_tp4_profile", "torch_profiler_record_shapes": "True"}' \
    --no-enable-prefix-caching

before：

after:

2.input rmsnorm and quant fused for kimi2.5
before：

after：

Test Result

accuracy:

Submission Checklist

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

[valarLip]

more clean up needed..

[qichu-yun]

Moved the definitions of self.quant_dtype and self.input_norm_quant_type into init to centralize initialization.