Drops redundant softmax unscale

Removes the unused reverse scaling parameter from the forward configuration to avoid stale values when softcap toggles.

Author: LoserCheems

csrc/flash_dmattn/flash_api.cpp

@@ -132,13 +132,11 @@ void set_params_fprop(
         params.softcap = softmax_scale / softcap;
         params.scale_softmax = softcap;
         params.scale_softmax_log2 = softcap * M_LOG2E;
-        params.unscale_softmax = 1.0f / softmax_scale;
     } else{
         // Remove potential NaN
         params.softcap = 0.0;
         params.scale_softmax = softmax_scale;
         params.scale_softmax_log2 = softmax_scale * M_LOG2E;
-        params.unscale_softmax = 1.0f / softmax_scale;
     }
 
     params.is_causal = is_causal;

csrc/flash_dmattn/src/flash.h

@@ -101,7 +101,6 @@ struct Flash_fwd_params : public QKV_params, public Mask_params, public Bias_par
     // The scaling factors for the kernel.
     float scale_softmax;
     float scale_softmax_log2;
-    float unscale_softmax;
     float softcap;
 
     // array of length b+1 holding starting offset of each sequence.