Remove pipeline/sdpa_patch.py (replaced by the flashrefocus package)

pipeline/sdpa_patch.py

@@ -1,28 +0,0 @@
-"""Route the Depth Anything V2 (DINOv2) attention through torch.nn.functional.scaled_dot_product_attention (SDPA).
-
-apply() replaces Attention.forward at runtime, without editing the submodule: plain softmax attention -> SDPA
-(Flash / memory-efficient kernels), 4.2x faster at 1536x1536 with the same result up to floating-point reordering.
-The renderer keeps its own attention: its short axial rows with a float bias run slower through SDPA."""
-import sys
-from pathlib import Path
-
-import torch
-import torch.nn.functional as F
-
-sys.path.insert(0, str(Path(__file__).resolve().parents[1]))  # repository root
-import paths  # noqa: E402
-
-
-def _da2_forward(self, x, attn_bias=None):
-    assert attn_bias is None
-    B, N, C = x.shape
-    q, k, v = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
-    x = F.scaled_dot_product_attention(q, k, v, scale=self.scale).transpose(1, 2).reshape(B, N, C)
-    return self.proj_drop(self.proj(x))
-
-
-def apply():
-    paths.add_to_path()
-    from depth_anything_v2.dinov2_layers import attention as da2_attn
-    da2_attn.Attention.forward = _da2_forward
-    da2_attn.MemEffAttention.forward = _da2_forward