drop the SDPA backend override: it made no difference to peak memory

app.py

@@ -29,14 +29,11 @@ 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")
+# The SDPA backend order that wan21_patches/modules/attention.py pins (cuDNN first)
+# is left alone. Forcing EFFICIENT_ATTENTION here was tried, on the theory that this
+# sm_120 / torch 2.11 runtime was falling through to the MATH backend; it changed the
+# peak not at all (44.2 GB either way), so the backend is not the problem and the
+# project's own measured order stands.
 
 import queue
 import shutil