feat: enable SwiGluQuant fusion for Qwen3 MLP

[nie-linfeng]

Description

Enable the DequantSwigluQuant fused ACLNN operator for Qwen3 decoder MLP across prefill and decode paths. This fuses SwiGLU activation and quantization into a single kernel, eliminating intermediate BF16 tensor round-trips between SwiGluForwardKernel and ElewiseQuant.

Performance benefits

Single operator profit of about 3us, whole network quantification of 30 layers-down proj profit of about 0.1ms

[gemini-code-assist]

Code Review

This pull request updates the Qwen3 decoder loader to cast the projection offset and scale tensors to float32, and enables SwiGLU quantization in the NPU decoder layer implementation. The reviewer noted a risk of an uninitialized member variable where enableSwigluQuant is not explicitly set to false when quantization is disabled, which could lead to undefined behavior, and suggested initializing it to false by default.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

Uninitialized Member Variable Risk

In C++, external library structs like atb_speed::qwen::QwenLayerParam may not default-initialize all of their member variables (such as enableSwigluQuant) to false in their default constructors.

If quantize_type_ is not "w8a8", or if down_proj_quantized() is false, enableSwigluQuant is never explicitly set to false in init_layer(). This can leave the field with garbage/random memory values, potentially enabling the SwiGluQuant fusion unexpectedly and causing runtime crashes or incorrect outputs.

To prevent this undefined behavior, please explicitly initialize enableSwigluQuant to false in param_from_args (or initialize_quantization_parameters):

void NpuQwen3DecoderLayerImpl::param_from_args(
    atb_speed::qwen::QwenLayerParam& param,
    const ModelArgs& args,
    const ParallelArgs& parallel_args,
    bool isPrefill) {
  // ... existing initializations ...
  param.enableSwigluQuant = false;
  // ...
}

[nie-linfeng]

fixed

[xiao-yu-chen]

please attach test performance benefits.