Why introduce the design of an `output_quantizer` within quantization operators?

I'm puzzled by a specific question: why does modelopt introduce the design of an output_quantizer within QuantOP(QuantInputBase, QuantLinearConvBase), then disable output_quantizer？(code from here)

I think there might be two places that could be the cause：

1. For kv cache quantization:

• run_auto_quantize.py set "*output_quantizer" to enable.(code from here)
• quantization/algorithms.py set "*output_quantizer" to false to diable kv cache quantization. (code from here)

However it now appears that configuring KV cache quantization is achieved via "*[kv]_bmm_quantizer".

2. For LayerNorm output, _FP8_MHA_OVERRIDE set "*output_quantizer" to enable(code from here). I am unclear on what "fuse the shared Q/DQ across all downstream Q/K/V/FC consumers" implies. Does this mean that the Q, K, V, and FC inputs are all expected to be FP8-quantized?

Are there other reasons related to the definition of output_quantizer that I am unaware of? Looking forward to your reply, Thanks in advance!

Start with modelopt/torch/quantization/nn/modules/quant_module.py and compare the output_quantizer handling with the referenced examples/llm_autodeploy/run_auto_quantize.py, modelopt/torch/quantization/algorithms.py, and examples/torch_onnx/torch_quant_to_onnx.py. Trace the KV-cache and FP8 MHA configurations, then document the rationale and clarify what behavior is expected when output_quantizer is enabled or disabled.