force the memory-efficient SDPA backend on sm_120

app.py

@@ -29,6 +29,15 @@ import os
 # allocator. Set it explicitly so a platform default cannot turn it back on.
 os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:False"
 
+# wan21_patches/modules/attention.py pins SDPA's backend order with cuDNN first,
+# from timings taken on an sm_100 B200 with torch 2.13. This Space is sm_120 on
+# torch 2.11, where that order falls through to the MATH backend: it materialises
+# the whole [B, 12, 2760, 29440] score matrix, and one block forward peaks at
+# 44.2 GB against a 47.4 GB slice. The memory-efficient fused kernel handles both
+# shapes this model attends over -- unmasked self-attention against the K/V cache,
+# and key-padded cross-attention -- in linear memory.
+os.environ.setdefault("WAN_SDPA_BACKEND", "EFFICIENT_ATTENTION")
+
 import queue
 import shutil
 import subprocess