From da35397017dbb16d71c6eddb4c3e1e0e4939bec6 Mon Sep 17 00:00:00 2001
From: winskuo-quic <winskuo@qti.qualcomm.com>
Date: Mon, 10 Nov 2025 16:09:39 +0800
Subject: [PATCH] Qualcomm AI Engine Direct - VIT Optimization

---
 examples/qualcomm/scripts/torchvision_vit.py | 74 +++++++++++++++++---
 1 file changed, 64 insertions(+), 10 deletions(-)

diff --git a/examples/qualcomm/scripts/torchvision_vit.py b/examples/qualcomm/scripts/torchvision_vit.py
index 2a428683ec3..ed8dbb792c4 100755
--- a/examples/qualcomm/scripts/torchvision_vit.py
+++ b/examples/qualcomm/scripts/torchvision_vit.py
@@ -7,12 +7,14 @@
 import json
 import logging
 import os
+from contextlib import contextmanager
 
 from multiprocessing.connection import Client
 
 import numpy as np
 
 import torch
+import torch.nn.functional as F
 from executorch.backends.qualcomm.quantizer.quantizer import QuantDtype
 from executorch.examples.models.torchvision_vit.model import TorchVisionViTModel
 from executorch.examples.qualcomm.utils import (
@@ -25,6 +27,56 @@
 )
 
 
+# Copied from torch/nn/functional.py
+# QNN does not have 5D permute optimization. Fuse to a single 4D optimization
+# Changed unsqueeze(0).transpose(0, -2).squeeze(-2) to permute(2, 0, 1, 3)
+def _in_projection_packed_custom(q, k, v, w, b=None) -> list[torch.Tensor]:
+    from torch.nn.functional import linear
+
+    E = q.size(-1)
+    if k is v:
+        if q is k:
+            # self-attention
+            proj = linear(q, w, b)
+            # reshape to 3, E and not E, 3 is deliberate for better memory coalescing and keeping same order as chunk()
+            proj = proj.unflatten(-1, (3, E)).permute(2, 0, 1, 3).contiguous()
+            # pyrefly: ignore  # bad-return
+            return proj[0], proj[1], proj[2]
+        else:
+            # encoder-decoder attention
+            w_q, w_kv = w.split([E, E * 2])
+            if b is None:
+                b_q = b_kv = None
+            else:
+                b_q, b_kv = b.split([E, E * 2])
+            q_proj = linear(q, w_q, b_q)
+            kv_proj = linear(k, w_kv, b_kv)
+            # reshape to 2, E and not E, 2 is deliberate for better memory coalescing and keeping same order as chunk()
+            kv_proj = kv_proj.unflatten(-1, (2, E)).permute(2, 0, 1, 3).contiguous()
+            # pyrefly: ignore  # bad-return
+            return (q_proj, kv_proj[0], kv_proj[1])
+    else:
+        w_q, w_k, w_v = w.chunk(3)
+        if b is None:
+            b_q = b_k = b_v = None
+        else:
+            b_q, b_k, b_v = b.chunk(3)
+        # pyrefly: ignore  # bad-return
+        return linear(q, w_q, b_q), linear(k, w_k, b_k), linear(v, w_v, b_v)
+
+
+# Context manager to patch temporarily, so it won't affect other users using F._in_projection_packed
+@contextmanager
+def PermuteInProjectionPacked():
+    # Save the original function so it can be restored later
+    _original_in_projection_packed = F._in_projection_packed
+    F._in_projection_packed = _in_projection_packed_custom
+    try:
+        yield
+    finally:
+        F._in_projection_packed = _original_in_projection_packed
+
+
 def main(args):
     # ensure the working directory exist.
     os.makedirs(args.artifact, exist_ok=True)
@@ -44,16 +96,18 @@ def main(args):
         )
 
     pte_filename = "vit_qnn_q8"
-    instance = TorchVisionViTModel()
-    build_executorch_binary(
-        instance.get_eager_model().eval(),
-        instance.get_example_inputs(),
-        args.model,
-        f"{args.artifact}/{pte_filename}",
-        inputs,
-        quant_dtype=QuantDtype.use_8a8w,
-        shared_buffer=args.shared_buffer,
-    )
+    instance = TorchVisionViTModel().get_eager_model().eval()
+
+    with PermuteInProjectionPacked():
+        build_executorch_binary(
+            instance,
+            inputs[0],
+            args.model,
+            f"{args.artifact}/{pte_filename}",
+            inputs,
+            quant_dtype=QuantDtype.use_8a8w,
+            shared_buffer=args.shared_buffer,
+        )
 
     if args.compile_only:
         return