Use torch swiglu

modeling_flashppi.py

@@ -12,7 +12,6 @@ from .configuration_flashppi import FlashPPIConfig
 
 # Detect Flash Attention installation
 try:
-    from flash_attn.ops.activations import swiglu
     from flash_attn.layers.rotary import apply_rotary_emb_func
     from flash_attn import flash_attn_varlen_kvpacked_func
     from flash_attn.bert_padding import pad_input, unpad_input
@@ -22,8 +21,8 @@ except ImportError:
     unpad_input = pad_input = apply_rotary_emb_func = None
     flash_attn_varlen_kvpacked_func = None
 
-    def swiglu(x, y):
-        return F.silu(x) * y
+def swiglu(x, y):
+    return F.silu(x) * y
 
 class RMSNorm(nn.Module):
     """RMSNorm without variance_epsilon buffer for checkpoint compatibility."""