Alpha Compute

void trt_edgellm::kernel::computeFC1Alpha(

float const *actGs,

float const *weightGs,

float *alpha,

int32_t numLocalExperts,

cudaStream_t stream

)

Compute per-expert FC1 alpha for the grouped-GEMM FP32 epilogue.

Forward-scale contract: actGs and weightGs are forward-direction global SFs. The kernel writes alpha[i] = (*actGs) * weightGs[i] — exactly the scalar applied before the activation inside the FC1 kernel (out[m,n] = act(alpha[e(m)] * acc)).

Parameters:

• actGs – [1] float32 on device — forward-direction activation GS.

• weightGs – [L] float32 on device — forward-direction per-expert weight GS.

• alpha – [L] float32 on device (output).

• numLocalExperts – L — number of local experts.

• stream – CUDA stream.

void trt_edgellm::kernel::computeFC2Alpha(

float const *actGs,

float const *weightGs,

float *alpha,

int32_t numLocalExperts,

cudaStream_t stream

)

Compute per-expert FC2 alpha. Same shape as computeFC1Alpha; distinct symbol provided for clarity at callsites.