fix: sink bias domain correction — add attn_sink/scale to raw logits

The softmax scales by scale_log2 = scale * log2(e). Adding sink_val to
raw logits causes it to be scaled too. Fix: add sink_val/scale instead,
so after scaling: (sink_val/scale) * scale_log2 = sink_val * log2(e).
This correctly multiplies attention weights by exp(sink_val).

dsv4/kernels/attention/fmha.py

@@ -434,9 +434,12 @@ class FmhaKernel:
                     kt_offset = Int32(kt * 128)  # KV position offset for this tile
                     # D5c: Read sink bias once (same for all positions in this head).
                     # Define unconditionally for CuTeDSL scoping (used when n_comp > 0).
+                    # The bias must be added in the SCALED-LOG2 domain: attn_sink * log2(e).
+                    # But we add to the RAW logits before the scale_log2 multiply.
+                    # Raw correction: attn_sink / scale → after * scale_log2 → attn_sink * log2(e)
                     sink_val = Float32(0.0)
                     if const_expr(self.n_comp > 0):
-                        sink_val = mSinkBias[Int32(0)]
+                        sink_val = mSinkBias[Int32(0)] / Float32(self.scale_softmax)
                     for j0 in range(32):
                         for j1 in range(4):
                             coord = tTMEM_LOADcS[(j0, 0), j1, 0, 0]