add stable diffusion sample code for onnxwrapper

Author: WeidongFeng916

README.md

@@ -120,8 +120,8 @@ We provide [QAI AppBuilder Launcher](tools/launcher/), enabling you to experienc
 ### 4. ONNX2BIN
 [ONNX2BIN](./tools/convert/onnx2bin/) is a guide to help you convert the ONNX model format into the BIN format optimized for a specific platform.
 
-### 5. ONNXRT2QNNRT
-[ONNXRT2QNNRT](./tools/onnxrt2qnnrt/) is a wrapper to run onnx inference code with qnn model, which will switch to qnn runtime automatically.
+### 5. ONNXWRAPPER
+[ONNXWRAPPER](./tools/onnxwrapper/) is a wrapper to run onnx inference code with qnn model, which will switch to qnn runtime automatically.
 
 ## Models
 ### Model Hub

script/qai_appbuilder/onnxwrapper.py

@@ -726,6 +726,26 @@ def Inference(self, input_tensors: List[np.ndarray], input_names: Optional[List[
         return final_outs
 
 
+# -------------------- ORT enum shims --------------------
+class GraphOptimizationLevel:
+    """Compatibility shim for onnxruntime.GraphOptimizationLevel.
+    onnxruntime uses integer levels 0..3; our SessionOptions.graph_optimization_level
+    already stores an int, so exposing these constants makes existing scripts work.
+    """
+    ORT_DISABLE_ALL = 0
+    ORT_ENABLE_BASIC = 1
+    ORT_ENABLE_EXTENDED = 2
+    ORT_ENABLE_ALL = 3
+
+class ExecutionMode:
+    """Compatibility shim for onnxruntime.ExecutionMode."""
+    ORT_SEQUENTIAL = 0
+    ORT_PARALLEL = 1
+
+class SessionOptionsMode:
+    """Placeholder for any future mode enums; kept for forward compatibility."""
+    pass
+
 # -------------------- SessionOptions --------------------
 class SessionOptions:
     def __init__(self):
@@ -734,6 +754,8 @@ def __init__(self):
         self.enable_profiling = False
         self.optimized_model_filepath = ""
         self.graph_optimization_level = 1
+        self.enable_mem_pattern = True
+        self.execution_mode = ExecutionMode.ORT_SEQUENTIAL
         self.qnn_runtime = Runtime.HTP
         self.qnn_libs_dir = ""
         self.qnn_profiling_level = ProfilingLevel.OFF

tools/onnxwrapper/Samples/stable_diffusion_1_5_onnx/README.md

@@ -0,0 +1,24 @@
+# stable_diffusion_v1_5 Sample Code
+
+## Introduction
+This is sample code for using AppBuilder to load Stable Diffusion 1.5 QNN models to HTP and execute inference to generate image.
+
+## Run the sample code
+```
+If you want to run the sample code with onnx models.
+python prepeare_stable_diffusion_onnx_models.py
+python stable_diffusion_1_5_onnx_infer.py --model_root models-onnx\modularai_stable-diffusion-1-5-onnx --provider cpu --out sd15_out.png
+
+If you want to run the sample code with qnn models.
+python prepeare_stable_diffusion_qnn_models.py
+python onnxexec.py stable_diffusion_1_5_onnx_infer.py --model_root models-qnn --vae_scale 1.0
+
+You also can add the following code at beginning of stable_diffusion_1_5_onnx_infer.py.
+from qai_appbuilder import onnxwrapper
+Then run the following command.
+python stable_diffusion_1_5_onnx_infer.py --model_root models-qnn --vae_scale 1.0
+
+```
+## Output
+The output image will be saved to sd15_out.png
+

tools/onnxwrapper/Samples/stable_diffusion_1_5_onnx/prepeare_stable_diffusion_onnx_models.py

@@ -0,0 +1,149 @@
+#=============================================================================
+#
+# Copyright (c) 2023, Qualcomm Innovation Center, Inc. All rights reserved.
+#
+# SPDX-License-Identifier: BSD-3-Clause
+#
+#=============================================================================
+
+# ---------------------------------------------------------------------
+# Download Stable Diffusion 1.5 ONNX models (with optional external data)
+#
+# Requirements:
+#   pip install -U huggingface_hub onnx
+# ---------------------------------------------------------------------
+
+import argparse
+from pathlib import Path
+
+from huggingface_hub import snapshot_download
+
+
+REPO_ID = "modularai/stable-diffusion-1.5-onnx"
+
+
+def _model_requires_external_data(onnx_path: Path) -> tuple[bool, set[str]]:
+    """
+    Return (requires_external_data, locations)
+    locations: set of external data filenames referenced by the model.
+    """
+    import onnx
+    from onnx import TensorProto
+
+    # Load without external data (we only need the metadata that says whether it uses external data)
+    m = onnx.load(str(onnx_path), load_external_data=False)
+
+    referenced = set()
+
+    def collect_from_tensor(t):
+        # TensorProto has data_location + external_data entries when external
+        if isinstance(t, TensorProto) and t.data_location == TensorProto.EXTERNAL:
+            # external_data is a repeated key/value pair list
+            loc = None
+            for kv in t.external_data:
+                if kv.key == "location":
+                    loc = kv.value
+                    break
+            if loc:
+                referenced.add(loc)
+
+    # Check initializers
+    for init in m.graph.initializer:
+        collect_from_tensor(init)
+
+    # Check constant nodes (rare but possible)
+    for node in m.graph.node:
+        if node.op_type == "Constant":
+            for attr in node.attribute:
+                if attr.type == onnx.AttributeProto.TENSOR:
+                    collect_from_tensor(attr.t)
+
+    return (len(referenced) > 0), referenced
+
+
+def download_onnx_models(out_root: Path):
+    print(f"[HF] Downloading ONNX models from: {REPO_ID}")
+    print(f"[HF] Output directory: {out_root}")
+
+    # Download ONNX + external data files (if they exist)
+    allow_patterns = [
+        "**/*.onnx",
+        "**/*.onnx_data",
+        "**/*.json",
+        "**/*.txt",
+        "**/*.md",
+    ]
+
+    snapshot_download(
+        repo_id=REPO_ID,
+        local_dir=str(out_root),
+        local_dir_use_symlinks=False,
+        allow_patterns=allow_patterns,
+    )
+    print("[HF] Download completed")
+
+    # ------------------------------------------------------------
+    # Post-check: only require external data when ONNX references it
+    # ------------------------------------------------------------
+    print("[CHECK] Verifying ONNX external data references...")
+
+    missing = []
+    checked = 0
+
+    for onnx_path in sorted(out_root.rglob("*.onnx")):
+        checked += 1
+        requires_ext, locations = _model_requires_external_data(onnx_path)
+
+        if not requires_ext:
+            print(f"[OK] {onnx_path.relative_to(out_root)} (no external data)")
+            continue
+
+        # If the model references external data, ensure all referenced files exist
+        all_ok = True
+        for loc in locations:
+            data_path = (onnx_path.parent / loc).resolve()
+            if not data_path.exists():
+                all_ok = False
+                missing.append(str(data_path))
+        if all_ok:
+            locs = ", ".join(sorted(locations))
+            print(f"[OK] {onnx_path.relative_to(out_root)} (external data: {locs})")
+        else:
+            print(f"[ERROR] {onnx_path.relative_to(out_root)} references missing external data: {locations}")
+
+    print(f"[CHECK] Scanned {checked} ONNX files.")
+    if missing:
+        raise RuntimeError(
+            "Some ONNX models reference external data files that are missing:\n"
+            + "\n".join(missing)
+            + "\n\nThis is required by ONNX external data format. "
+              "Ensure the referenced files exist in the same folder as model.onnx. "
+              "See ONNX external data spec for details."
+        )
+
+    print("[CHECK] All required external data files are present ✅")
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Download Stable Diffusion 1.5 ONNX models (with optional external data)"
+    )
+    parser.add_argument(
+        "--out_dir",
+        type=str,
+        default="models-onnx/modularai_stable-diffusion-1-5-onnx",
+        help="Output directory",
+    )
+    args = parser.parse_args()
+
+    out_root = Path(args.out_dir).resolve()
+    out_root.mkdir(parents=True, exist_ok=True)
+
+    download_onnx_models(out_root)
+
+    print("\n[OK] ONNX model preparation finished.")
+    print(f"Model root: {out_root}")
+
+
+if __name__ == "__main__":
+    main()