[Bugfix] Fix logit soft cap in flash-attn backend (vllm-project#7425)

vllm/attention/backends/flash_attn.py

@@ -563,6 +563,7 @@ def forward(
                 softmax_scale=self.scale,
                 causal=True,
                 alibi_slopes=self.alibi_slopes,
+                softcap=self.logits_soft_cap,
             ).squeeze(1)
 
         # Reshape the output tensor.