Drop all-True attention mask -> flash SDPA (memory fits on MIG)

transformer_krea2.py

@@ -525,6 +525,11 @@ class Krea2Transformer2DModel(ModelMixin, ConfigMixin, AttentionMixin, PeftAdapt
             ref_temb_mod = self.time_mod_proj(F.gelu(temb_zero, approximate="tanh"))
             block_temb = (temb_mod, ref_temb_mod, text_seq_len + image_seq_len - ref_seq_len)
 
+        # An all-True mask is equivalent to no mask; passing None keeps SDPA on its fast, low-memory (flash)
+        # path instead of the mask-materializing math fallback — critical on small / MIG GPUs.
+        if encoder_attention_mask is not None and bool(encoder_attention_mask.all()):
+            encoder_attention_mask = None
+
         text_attention_mask = None
         attention_mask = None
         if encoder_attention_mask is not None: